AI Sales Co-Workers:
What They Are, How They Work, and How to Get Started Today

What is an AI sales co-worker? How it compares to AI sales assistants, agents, and tools

The terminology is a mess. Vendors call their products AI sales agents, AI sales assistants, AI sales colleagues, and AI execution platforms, sometimes interchangeably, sometimes to describe fundamentally different things. Before we go any further, let's get the definitions straight, because they matter more than most people realise.

The spectrum from chatbot to co-worker

There's a clear progression in how AI tools relate to your sales workflow:

The defining characteristic of a co-worker model is proactive execution. No login required. No prompt needed. It monitors what's happening across your entire pipeline, identifies what needs attention, and tells you, or does something about it directly.

What it's not

An AI sales co-worker is not another dashboard. It is not another analytics layer. It is not another tool that requires reps to change their behaviour, adopt a new interface, or remember to check it. If it requires effort from the rep to generate value, it's an assistant at best. A co-worker delivers value into the rep's existing day; they don't have to come to it. If you have to log in to see the value, it isn't a co-worker.

How it works at a high level

An AI sales co-worker ingests your CRM data, email threads, call transcripts, and meeting notes continuously. It builds a structured understanding of every open deal: who the stakeholders are, where things stand, what's been said, what's been promised, and what's been conspicuously quiet. Then it acts on that understanding: surfacing risks, drafting communications, flagging stalled deals, preparing reps for upcoming calls, and writing CRM updates back automatically.

The best proof point is also the simplest to picture: before any rep has opened their CRM in the morning, an AI co-worker already knows which three deals need attention today and why, and has already drafted what to do about each one.

Where in the sales cycle does an AI co-worker sit?

Most of the AI sales tooling built in the last two years targets the same problem: generating pipeline. AI-SDRs automate prospecting, personalise outreach at scale, and book meetings. That is genuinely useful for teams struggling with top-of-funnel volume, and the category has attracted significant investment and adoption.

An AI sales co-worker solves a different problem. It operates after a deal enters the pipeline: managing the complexity of active deals, keeping multi-stakeholder relationships moving, ensuring follow-through happens, and reducing the gap between identifying a risk and acting on it. The handoff point, roughly, is the first qualified conversation. That same logic extends beyond the close: the relationships, context, and signals built during the deal cycle are exactly what an AI co-worker uses to surface expansion opportunities and flag renewal risk before it becomes a problem.

If pipeline creation is the constraint, an AI-SDR is the right category to evaluate. If the problem is that deals stall, contacts go dark, business cases never get written, and reps are triaging rather than executing, that is where an AI co-worker operates.

This is not a minor distinction. The largest individual deals, where a single closed-won represents months of quota attainment, are almost never lost at the top of the funnel. They are lost in the middle: a stakeholder who disengaged, a follow-up that never happened, a business case that never reached the CFO. That is the gap an AI co-worker is built to close.

Beyond AI sales automation: why co-workers execute instead of just automate

The past three years in B2B sales technology have been dominated by one question: how do we get better visibility into the pipeline? The answer was a wave of analytics, conversation intelligence, and revenue forecasting tools. Platforms like Gong and Clari built strong businesses on this insight, and they're genuinely useful: they show you what happened, surface patterns across deals, and flag where things went wrong.

But here's the problem. Knowing what went wrong in a deal and fixing it are two completely different capabilities. The game film is valuable. Someone still has to play the game.

The Monday manager review

Picture the most common scene in sales leadership: a Monday pipeline review. The manager works through open deals, asks where things stand, and surfaces a dozen insights. This deal hasn't had contact in three weeks. That one is missing key stakeholders. This one has the right activities but no documented business case. The forecast risk is clear.

And then what? The manager has 45 minutes to cover 40 deals. By Wednesday, they're back in their own calls. By Thursday, the insights from Monday are stale. The rep who needed to re-engage that quiet deal got busy with something else. Nothing changed.

This is the insight-to-action gap. It's not a failure of analysis; it's a failure of execution bandwidth. Most sales organisations today have more data than they can act on. The bottleneck isn't knowledge. It's the capacity to do something with it.

Why adding more intelligence doesn't close the gap

The instinct is to add more tools: better forecasting, more granular deal scoring, AI-generated recommendations. And these help, up to a point. But they all have the same structural problem: they produce outputs that humans still have to act on.

The insight-to-action gap doesn't close by making the insights sharper. It closes when you add an execution layer: something that doesn't just surface what needs to happen, but actually starts making it happen. That's the category shift the AI co-worker represents. Without it, organisations accumulate what you might call execution debt: insights that were surfaced, understood, and then never acted on because nobody had the bandwidth.

Gartner projects that 75% of B2B sales organisations will augment their traditional playbooks with AI-guided selling by 2025.² The gap between those who add an execution layer and those who stop at visibility is where the category separates.

The real cost of manual selling, and what AI for sales teams actually changes

Here's a number worth sitting with: sales reps spend just 28%¹ of their time actually selling. The rest goes to what we call Shadow Work: CRM administration, internal communication, building business cases, writing follow-ups, preparing for calls, and all the coordination overhead that accumulates in a complex deal cycle. Break it down further: 17% of the average week goes to CRM data entry alone, and another 14% to account research and call preparation.¹

This is not a performance problem. The reps doing this aren't lazy or disorganised; they're doing exactly what the job requires. The problem is that nobody designed the system around the question of how much of this work could be handled differently. It's just always been this way.

What a typical sales morning looks like

It's 9am. Before a rep has their first prospect conversation, here's what typically happens: 40–45 minutes in the CRM, reviewing open deals, updating stages, figuring out what needs attention today. Then 30 minutes writing a follow-up to a discovery call from last week, pulling together what was discussed, what was promised, what the next step is. Then 20 minutes building a one-pager for a champion who needs to make an internal case for budget. That's 90 minutes of overhead before the day's selling even starts.

Multiply that across a team of 20 reps, 250 selling days a year, and a loaded cost of €80k per head, and the arithmetic gets uncomfortable quickly.

The silent deal killer

Admin overhead is visible and measurable. The silent deal killer is harder to quantify, but more damaging: deals that go cold not because the rep gave up, but because they had 29 other deals and this one fell through the cracks.

No rep is ignoring deals deliberately. They're triaging. With limited attention and a full pipeline, some deals inevitably get less follow-through than they should. The prospects on the receiving end of that silence don't know it's a capacity problem; they just know the vendor went quiet. And they move on.

The 80/20 problem, and its real cause

Typically, around 20% of reps generate 80% of revenue — sometimes the concentration is even tighter than that. The instinctive explanation is talent: some reps are just better. And that's partly true. But the more accurate explanation is systems. Top performers aren't working harder. They're more systematic: more consistent with follow-up, more thorough with meeting prep, more disciplined about CRM hygiene. They've built habits that the rest of the team hasn't.

The question an AI co-worker answers is: what if every rep had those same habits institutionalised for them? Not trained; institutionalised. The system does the follow-up consistently, the meeting prep thoroughly, the CRM update automatically. The top performer's approach, applied to every rep on the team. You aren't just buying software; you're upgrading the operating system of your entire sales floor.

What an AI sales co-worker does day-to-day: deal intelligence, actions, and execution

Enough framing. Let's make this concrete. Here's what an AI co-worker looks like inside a sales team on an ordinary Tuesday.

That's the scenario that tends to make sales leaders stop and say “I want that.” Everything else is supporting evidence for why it's possible.

Five more scenarios that show the range

The common thread across all of these: the co-worker knows the deal, knows the history, knows what's missing, and acts without waiting to be asked. That's the distinction that matters.

The business case that writes itself

Gong's analysis of over 1.8 million deals found that closed-won deals have twice as many buyer contacts as lost deals, and in enterprise transactions, won deals average 17 stakeholders involved.³ The implication is direct: winning isn't just about convincing one person. It's about enabling the people inside the buying organisation to convince each other. That's buyer enablement, and it's where most deals are actually won or lost.

One of the more consequential moments in any complex deal is when the champion needs to justify the purchase internally: to their CFO, their procurement team, or a steering committee they don't fully control. Most deals slow down or die here, not because the value isn't there, but because nobody has turned it into numbers the finance team will accept.

Optivian generates a full business case document automatically once the deal has the right context: an identified economic buyer, a confirmed deal amount, documented decision criteria, and quantified metrics from the buyer's own conversations. When those conditions are met, the document appears in the deal, with no prompt from the rep required. It includes projected financial gains broken out by value driver, a full ROI and payback period analysis, and a methodology section that ties every projection back to real data: the buyer's stated numbers, your existing case studies, and industry benchmarks. Projections are marked as estimates throughout, which keeps the document credible rather than promotional.

The rep can also regenerate it with specific instructions: “focus on cost savings, their CFO doesn't care about time savings” or “use more conservative projections, their finance team will push back.” The document updates accordingly. The champion walks into the budget conversation with something concrete. The rep didn't spend an afternoon building it.

Is your team ready for an AI co-worker? What AI for sales teams actually requires

The most common reason sales orgs delay AI adoption is the belief that their data isn't good enough to start. “Our CRM is a mess.” “We don't have call recordings.” “We haven't been logging consistently.” We hear this constantly, and it almost never means what teams think it means.

What you actually need to start

The minimum viable setup is simpler than most teams expect:

• ✦A connected CRM that acts as a system of record, identifies who owns the opportunities and what the close dates are. That's it. Everything else gets built from the interaction data.
• ✦Emails logged in CRM (which most teams already have, since modern CRMs do this automatically).
• ✦Call transcripts: a meeting recorder handles this. No manual effort required.
• ✦Notes for physical or phone conversations: the one place where some logging discipline is genuinely required. But a short voice note after a call is all it takes.

That's it. No historical data requirements. No minimum deal count threshold. An AI co-worker can start adding value on day one of a deal, beginning from zero.

What about messy CRM data?

Here's the part that surprises most teams: the AI co-worker is often most valuable to orgs with imperfect data hygiene. Why? Because its first job is to surface exactly where the gaps are. It reads what's there, infers what's missing from the email and call data, and flags the discrepancies. Teams frequently discover that their “messy CRM” problem is actually more of a visibility problem; they didn't know which specific fields were unreliable or inconsistent until the AI started surfacing the gaps deal by deal.

What makes it better over time

The system gets meaningfully smarter as it accumulates context: more interaction data, more complete stakeholder coverage, more deal history to learn from. But this is an upside, not a prerequisite. You don't need to reach a certain level of data quality before you start; you start, and the data quality improves as a by-product of the system running.

Who this isn't for

No interaction data. An AI co-worker reads the record of your deals: emails, calls, meeting notes. If none of that exists, there's nothing to work from. That's the only hard prerequisite.

High-volume, transactional sales. If deals close on a single call and ACV is low, this model is over-engineered for the job. It earns its value in complex sales with multiple stakeholders and longer cycles.

One clarification on scope: an AI co-worker is not an AI SDR. It doesn't create pipeline. It works on the pipeline you already have, helping you win deals in play and protect accounts already closed.

Build vs. buy: the AI in sales decision every revenue team has to make

In every enterprise sales process, someone on the technical or product side raises the same question: “Why can’t we just build this ourselves? We have engineers. We have Claude Code, Lovable, all the vibe-coding tools. This can’t be that hard.”

It’s a reasonable question. And to be fair: the prototype is easy. Here’s the honest answer about what comes after: the AI model is the easy part.

What ChatGPT actually does for sales, and where it stops

ChatGPT is a genuinely impressive writing tool. It can draft emails, summarise meeting notes, and generate talking points, given the right inputs. Tools like Claude Code or Lovable can take that further: you can spin up a working internal app in an afternoon. The limitation is everything before “given the right inputs.” None of these tools have memory of your deals. They don’t know that the prospect at Acme Corp mentioned budget constraints in week two, or that the champion went quiet after the security review, or that the economic buyer has a Q2 deadline. Every prompt starts from scratch. The output is only as good as what you put in, and building the context layer that fills the prompts automatically, continuously, across every deal in your pipeline, is where the real engineering work is.

The three hard problems that vendors have already solved

Building a real AI sales co-worker requires three capabilities that are each non-trivial:

• The context engine: continuous, bidirectional CRM sync; email, call, and meeting ingestion; structured profile-building for every deal and stakeholder. This is ongoing infrastructure work, not a one-time build.
• The proactive execution layer: event monitoring across all deals simultaneously, trigger logic that knows when to surface something and when not to, intelligent delivery into Slack, Teams, or email at the right moment. This is where most internal builds fail, not because the AI is wrong, but because the scheduling and delivery layer is harder than expected.
• The integration layer: CRM writeback, Slack and Teams delivery, calendar integration, voice of the rep personalisation. Each integration is a maintenance surface that requires ongoing attention as APIs change.

The realistic build cost

A realistic internal build takes 6–12 months of engineering time to reach feature parity with a mature product, and that’s before accounting for ongoing maintenance, model updates, and the iteration work required to make the outputs actually useful rather than technically correct. The teams we’ve seen attempt this most seriously hit the same wall: they can surface insights, but they can’t turn them into proactive, personalised execution at scale. The last mile is the hardest mile.

The build question is ultimately a focus question. The teams that try to build this internally spend engineering cycles on sales infrastructure instead of their core product. The teams that buy spend a fraction of the cost and start seeing results in weeks, not quarters.

How to make the business case for Digital Labor

CFOs and procurement teams would often describe AI co-workers as Digital Labor. It’s how analysts classify AI that works as productive workforce capacity rather than just another software subscription. That framing can help, because it moves the conversation away from tool costs and towards a workforce investment. Here’s the framework for making that case.

Start with the cost of the current state

The ROI question isn’t “what does this cost.” It’s “what does manual selling cost right now.” Don’t frame this as new software spend. Frame it as reclaiming lost capacity. Most business case conversations start in the wrong place: they try to justify new spend. The more effective framing is to quantify the existing cost and position the solution as the fix.

Start with the selling time math: if your team of 25 reps each earns €80k loaded cost, and they’re spending 65–70% of their time on non-selling work, you’re paying approximately €1.3m per year for activities that aren’t revenue-generating. That number is already on the books. You’re not asking for new budget; you’re asking for a better use of existing spend.

Sales velocity: the number that makes the business case

Sales velocity is a single metric that captures how much revenue your pipeline generates per day. It combines four variables: how many deals are open, their average value, how often you close them, and how long each cycle takes. Change any one variable and the number moves. Improve two simultaneously and the effect compounds.

The research makes clear why win rate is the highest-leverage variable, and why it’s so hard to move without arming the champion. Gartner finds that buyers spend just 17% of their total buying journey in direct conversation with potential suppliers.⁴ The other 83% happens in internal meetings, email threads, and budget conversations where no rep is present. The pitch isn’t what closes the deal. What closes the deal is what the champion says when they’re presenting to their CFO alone.

The execution gap is just as stark. RAIN Group research finds that top-performing sellers are 63% more likely to communicate a strong ROI case than their peers. Yet only 16% of buyers say their sellers are actually effective at making it.⁵ The business case is the most high-leverage document in a complex deal, and the one most reps never get around to building.

An AI co-worker directly affects two of the four velocity variables: win rate (through better-prepared reps, tighter qualification, and a business case that actually reaches the CFO) and cycle length (through consistent follow-up, faster re-engagement, and better internal mobilisation at the buyer’s end). The compounding effect of improving both simultaneously is where the business case gets genuinely interesting. Use the calculator below to see what it looks like with your own pipeline numbers.

The headcount replacement angle

For some organisations, the clearest ROI calculation isn’t about selling time or win rates; it’s about headcount. We’ve worked with enterprise sales teams that were seriously considering hiring 4–5 people to write business cases for their sales team, at a cost of €80–120k per head loaded. An AI co-worker automates that work at a fraction of the cost, and at higher consistency and speed. Replacing a hire versus buying a software subscription is a fundamentally different financial conversation, and it typically wins easily on the numbers.

The Supermetrics data point

Supermetrics processes 15% of global digital advertising spend and provides marketing data connectivity to over 200,000 companies worldwide. Optivian users at Supermetrics showed a 20% higher win rate than their peer group within the same organisation: same product, same market, same management. The difference was the system, not the talent.

The three-line CFO version

• Cost of the problem: €[X]M per year in rep time spent on non-selling work, based on [N] reps × €[Y]k loaded cost × 65% non-selling time.
• Impact on sales velocity: Improving win rate by 10–15% and shortening cycle by 15–20 days increases daily revenue by approximately €[Z]k, based on current pipeline metrics.
• Alternative to headcount: For [use case], the alternative to this solution is hiring [N] people at €[X]k each. That’s €[Y]k annually in perpetuity, versus a software subscription.

EU buyers: compliance, privacy-by-design, and the AI veto risk

If you’re a European enterprise, there’s a step in every AI procurement process that doesn’t appear on most vendor websites: the legal and IT review. It doesn’t always block deals, but when it does, it blocks them completely, and the criteria that trigger it are specific and predictable.

The AI procurement veto

In EU enterprise sales organisations, legal or IT teams typically have veto power over new AI tools. The triggers are usually one or more of: GDPR concerns about what data the tool processes and where; data residency questions about where customer data is stored; EU AI Act applicability; and security documentation requirements (SOC 2, ISO 27001) that procurement processes now routinely demand.

The good news is that these questions are answerable, if the vendor has designed their architecture to answer them. The bad news is that many AI sales tools, particularly US-built products focused primarily on the North American market, haven’t.

What to ask any AI sales vendor about compliance

• Where is our data processed and stored? EU residency or explicit transfer safeguards should be the baseline requirement.
• What data does the product actually need? A read-heavy, minimal-permissions model is meaningfully lower risk than tools that require full CRM write access. Less data access is less risk. Ask what the minimum is.
• What security certifications do you hold? Understand what’s current and what’s on the roadmap; factor in the timeline if your procurement process has a minimum certification requirement.
• How do you handle the EU AI Act? For tools that support or influence commercial decisions, this is increasingly relevant. Get their position in writing.
• Is your security architecture documented for IT review? A vendor who can’t produce security documentation promptly is a vendor who hasn’t prioritised this, which tells you something about their enterprise readiness.

The principle that matters most

The distinction between security built-in from day one and security bolted on later is visible in the architecture. Vendors who built EU enterprise deployments in mind from the start have cleaner answers to all of the above. Vendors retrofitting compliance onto a US-first product tend to have more qualifications, more exceptions, and more “on the roadmap” responses. Ask the questions early. The answers are diagnostic.

How to evaluate AI sales co-workers, and what separates them from generic AI sales tools

Most AI sales tools are not AI co-workers. Many call themselves co-workers, colleagues, or agents, and deliver something closer to a sophisticated chatbot with a CRM integration. The evaluation criteria are different, and applying the wrong criteria leads to choosing the wrong thing.

The difference comes down to three independent factors. Every AI in a work context has some level of each. What separates a co-worker from a tool is how strong all three are simultaneously:

These are independent levers. An AI with Knowledge and Skills but no Ownership is a tool, not a co-worker. It waits to be asked. Here’s what to test and ask during any evaluation:

The 7-question checklist

• ✦Does it act proactively without prompting? The single most important question. If the rep has to open the tool and ask it something before it provides value, it’s an assistant, not a co-worker. Test this specifically: run it on a live pipeline for two weeks and count how many times it surfaced something the rep didn’t explicitly request.
• ✦Does it have persistent deal context, or does it start fresh each time? An AI that doesn’t remember what happened in the last call is not a co-worker; it’s a stateless tool. Ask the vendor specifically how deal memory works and how far back it goes.
• ✦Does it deliver into your existing channels, or does it require a new login? Tools that require reps to adopt a new interface add friction and reduce adoption. A co-worker should show up in Slack or Teams, in the CRM, or in email, where reps already live.
• ✦Does it write in the rep’s voice, or generic AI voice? Generic AI-drafted emails get lower response rates and feel off-brand. Ask for a side-by-side: a draft from the tool versus a draft the rep would write themselves. The gap should be small.
• ✦How does it handle EU compliance? See Section 08. For EU enterprise buyers, this isn’t optional. Get answers to the compliance questions before you get deep into a technical evaluation.
• ✦Does it get smarter over time? A system that treats every deal the same in month six as it did in month one isn’t learning from your business. Ask how the system improves with more data and feedback from the team.
• ✦Can you configure automation without an engineer? If every workflow change requires a developer, the tool will calcify around the first use case it was set up for. Sales processes change. The tool needs to change with them, at the pace a RevOps team can manage.

Meet Ollie, your new colleague

If you’ve been reading this guide and thinking “this is exactly what my team needs”: that’s Ollie.

Ollie is Optivian’s AI sales co-worker: built to monitor every deal in your pipeline continuously, surface what needs attention before anyone asks, draft the follow-ups and business cases and meeting preps your reps don’t have time to do manually, and write everything back to the CRM automatically. It keeps working after the close, flagging expansion signals and renewal risk in existing accounts before they become emergencies. It lives in Slack or Teams and your CRM. It doesn’t require a new login. It doesn’t ask your reps to change what they do; it changes what happens between the things they do.

Ollie is built on three layers that work together continuously:

It’s currently running inside several mid-market and enterprise B2B sales teams. In the first week of one recent deployment, there were more than 500 interactions with Ollie across Slack and the platform; none of them initiated by the customer. Ollie found things to surface. That’s what a co-worker does.

Getting started: what setup actually involves

Implementation complexity is a real objection, and a fair one. In practice, teams start seeing Ollie surface useful work within the first week of connecting their CRM. Here’s what it actually involves:

• CRM connection (admin, ~15 minutes): One OAuth connection to your CRM. No API keys, no manual data mapping. Optivian begins ingesting your pipeline immediately after authorisation.
• Company Profile (admin, a few hours): Write in your positioning, a handful of case studies, and your key sales rules. This is what grounds every AI output in your actual business. The richer it is, the better Ollie’s output from day one.
• Reps connect their accounts (~5 minutes per rep): Each rep links their email and optionally Slack or Teams from their profile settings. This is what enables Ollie to learn their personal writing style and deliver work via the tools they already use.
• Ollie learns each rep’s preferences (a short conversation): Type /onboard in any Ollie chat. Ollie asks about role, territory, and communication style, saving the answers as persistent preferences that carry through all future interactions.

Stakeholders needed: one admin or RevOps person for the CRM connection and Company Profile; reps handle the rest themselves. No IT project. No custom development. In our experience, the first useful Ollie output arrives within days of setup completing: a prioritised action plan, a drafted follow-up, a generated business case.

See Ollie working on your pipeline

Book a 30-minute demo. We'll run Ollie on a sample of your actual deals: no curated data, no polished scenario. You'll see what it surfaces and what it would do next.

Book a demo →

Or talk to Ollie directly. Ask it anything about what an AI co-worker can do for your team.

Recommended by

Supermetrics

Halon

Sellforte

Kaks IO Labs

Buddy Healthcare

Earned autonomy: how much can you trust an AI sales agent or co-worker?

There’s a question underneath almost every conversation we have with sales leaders that doesn’t always get asked directly: “I get it, but how much can I actually trust it to run without oversight?”

It’s the right question. And the honest answer is: it depends on how far you are in the relationship.

The trust curve

The way an AI co-worker earns autonomy mirrors the way a new hire does:

• Weeks 1–4 (observe and surface): Ollie reads the pipeline, builds context, and starts surfacing suggestions. Everything is visible to the rep. Nothing happens without explicit approval. The rep is evaluating the output quality: is this useful? Is it accurate? Does it understand our deals?
• Month 2–3 (act with review): Drafts are reviewed rather than written from scratch. CRM updates are approved in one click rather than built manually. The rep spends minutes on tasks that used to take 30. Trust is expanding because the output quality is consistent.
• Month 4–6 (proactive ownership): The system is trusted to act on certain tasks automatically (CRM updates, meeting summaries, routine follow-ups) while still surfacing the more complex decisions for human review. Autonomy has been earned through a track record of accurate, contextually appropriate output.

What “trust” actually means in practice

It doesn’t mean trusting the AI blindly. It means trusting the output enough to review rather than redo from scratch; the same relationship you’d have with a competent junior team member. You don’t rewrite their draft from zero. You read it, make edits where needed, and send. That’s a much faster loop than building it yourself, and the quality bar is achievable from early in the deployment.

Practical controls

Every element of Ollie’s autonomy is configurable. Reps can set their preference anywhere from “suggest only: I’ll decide everything” to “act and confirm: execute within defined parameters and flag the exceptions.” The default starting position is conservative: suggest, surface, draft. The rep is always in the loop. Expanding scope happens when the team chooses to expand it, not on a vendor’s timeline.

The most common adoption pattern: teams start in suggest-only mode, see the quality of outputs over the first month, and organically start treating Ollie like a trusted colleague rather than a tool to double-check. The expansion typically happens without a formal decision; it just becomes the way the team works.

Salesforce’s 2025 State of Sales research found that 94% of sales leaders who work with AI report it as essential to meeting business demands, with the consistent emphasis that AI works as an expansion of capability, not a replacement for the person using it. That framing matches what we see in practice.