The AI Transformation Loop: how to pick one workflow and prove it

What we keep hearing

They bought the stack. People are using it. Nothing is showing up in the business.

That’s 15 to 20 conversations in the last month, with mid-market companies, associations, and nonprofits early in their AI work, and it’s close to the same conversation every time.

Usually it started with a mandate. We need to be AI-forward. Everyone needs to be using AI. What that means exactly, nobody defined, so the organization did the reasonable thing and bought a platform license for everybody. Now there’s real usage. There’s also nothing anyone can point a CFO at.

What’s underneath it is almost always the same. The organization embraced AI as the next big technology initiative instead of starting from a defined business problem and working through the changes that problem demands. They’re governing AI like legacy IT. And they’re mistaking personal productivity gains for enterprise value, which are genuinely different things.

Buying was never the hard part. The stack is a purchase order. What you run on top of it, one workflow at a time, is the whole job.

This piece is how we approach that. Five lenses for finding the one workflow worth proving, and a seven-step loop for proving it. The loop takes four to five weeks and produces a decision, not a transformation. That distinction is where most of this goes wrong, so I’ll be specific about it.

One note on the numbers below. Everything here is sourced. Where a study was vendor-sponsored or run by someone selling the answer, I flag it once so you can weight it yourself.

What the data says about that

94% of mid-market companies are already using generative AI. 2% have operationalized it at scale.

That’s from a December 2025 survey of 100 senior decision-makers at US companies between $5 million and just under $1 billion in revenue, every one with direct authority over AI spend (advisory-commissioned). The spread between those two numbers is the entire problem.

That gap is the whole story, and it shows up again the moment you look at whether anyone opens the thing you bought them. An independent survey of more than 150,000 respondents, run between July 2025 and January 2026, measured what share of people with workplace access to a platform actually use it. ChatGPT converted at 83.1%. Copilot at 35.8%. Gemini at 34.0%. The firm titled the writeup “Why Licenses Don’t Equal Adoption.” In their larger sample of 193,266 people, 38% said their company has no formal AI initiative, or weren’t sure one existed.

Fine, you might say, that’s a tooling problem. Except the gains people do get don’t add up to anything the business can see either. The cleanest study on this links AI adoption surveys to administrative labor records across Denmark. It found “precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2% two years after.” Those nulls held for intensive users, for early adopters, for workplaces with substantial investments, and for workers who reported large gains themselves.

Zoom out to the P&L and it’s the same picture. Only 37% of organizations report AI contributing positively to EBIT at all, and that number didn’t move year over year.

So where does the saved time actually go? Three places, all measured.

Some of it comes back as rework. BetterUp Labs and Stanford’s Social Media Lab surveyed 1,150 US desk workers and found 40% had received “workslop” in the prior month, AI-generated output that looks finished and isn’t. Two hours on average to resolve each one. They put it at $186 per employee per month. Your colleague’s saved hour became your two.

Some of it gets absorbed. In an Upwork study of 2,500 people, 96% of C-suite leaders expected AI to boost productivity, while 77% of employees using AI said the tools had ADDED to their workload.

And some of it never surfaces, because people don’t report it. Ethan Mollick has documented this for years: employees hide AI use because of policy risk, because of how AI-assisted work gets perceived, and because automating your own job is a strange thing to advertise.

There are two different things called AI here

Most of the confusion resolves once you separate two things that share a name.

Tools are personal. Summarizing documents, brainstorming, drafting emails, taking meeting notes. They’re genuinely useful and they save a few minutes per task. Everyone curious in your organization has already found them.

Solutions operate at a different scale. A model providing real-time coaching to call center agents. Something wired into your processes and your systems of record. These need integration, which is far more complex than employing AI for personal use.

The early interest you’re seeing is almost entirely the first kind, fueled by individuals who want to work more efficiently. That’s real, and it’s where most organizations stop. Tools everywhere, a platform contract, no solution anywhere.

That’s the mechanism behind every number in the last section. Personal productivity does not aggregate on its own. Somebody has to build the thing that operates at the level the business measures.

Why nobody ever finds the workflow

The obvious answer is to go find a workflow worth doing properly. The reason that doesn’t happen is more mundane than most strategy decks admit.

These are businesses with clients, customers, or members to service. They have relationships to protect. And their customers aren’t asking for AI. Nobody’s members are writing in to request an agent.

So there’s no forcing function. Their week this week looks like their week last week, and this month looks like last month. To answer “what should we automate,” somebody has to stop running the business long enough to look at it, and nothing in their day makes that happen.

Attention is the real constraint. Not budget, talent, or tooling. Any approach that needs the organization to pause for a strategy offsite loses to the Tuesday it’s competing with.

So the lenses have to work without that pause. They run on what the business already has in front of it.

Five lenses for finding the one workflow worth proving

Whatever you do, get outside the building. Stay inside it and you’ll struggle to know where to start, because from in there every process looks equally like a process. Start with the customer, the client, the member, and work inward.

Lens 1: Signal. Does a customer actually feel this?

Talk to your customer-facing people and mine your customer data. You’re looking for a specific moment: where a customer, member, or client is waiting, complaining, confused, or leaving.

Notice what you’re not asking. You’re not asking where to cut costs. You’re asking how to make their experience with your business better. Different question, different list.

Onboarding and the early sales conversations are usually the richest ground, and there’s a world of data sitting there that people overlook. Where are they dropping? Why didn’t we win that deal?

For associations this is measurable in a way that should get your attention. The Membership Marketing Benchmarking Report, drawing on nearly 700 associations, found that organizations with declining membership are significantly more likely to have first-year renewal rates below 60%. First-year renewal runs about ten points under overall renewal as a rule. The first-year experience tracks whether the organization is growing or shrinking.

Nonprofits show the same shape, with new-donor retention falling at roughly an order of magnitude worse than repeat-donor retention.

One limit worth naming. The evidence on member and donor onboarding is solid. My claim that the sales-to-delivery handoff specifically is where revenue gets won or lost is an operating observation from our engagements, and I couldn’t find published research establishing it. Take it as experience.

Lens 2: Work. Which steps are deterministic, and which are judgment?

Take the process you identified and walk it. Sort every step: deterministic, judgment-heavy, exception-driven, or relationship-dependent.

Machines take the deterministic steps and prepare the judgment ones. People keep the decisions, the approvals, and the exceptions.

If most of the work turns out to be judgment or relationship, it’s a bad first candidate. Save it for later. That’s a statement about what you can prove in a month, not a permanent verdict on the workflow.

This is also where you find out how much of the process nobody wrote down, which brings us to the next one.

Lens 3: Context. Does the truth exist, and does it have an owner?

Find where the truth lives, who owns it, and what it means. Every dataset in the workflow needs an owner, probably a definition, a sensitivity level, and a permissions level before an agent ever sees it.

In our work, context is what decides whether an agent succeeds. We’ve never had one fail because the model wasn’t smart enough. They fail because nobody could tell the thing what was true. That’s our experience, not a measured finding, and I’ll say why that matters. There’s a widely repeated claim that some large percentage of agent success comes down to context. We went looking for the study behind it. There isn’t one. It traces to a blog post that cites nothing. So take the judgment and ignore the number.

There’s a real problem with this lens, so let me say it before you do. Gartner reports that 63% of organizations either don’t have, or aren’t sure they have, the right data management practices for AI, and that just 4% have AI-ready data. Apply this lens strictly and almost nothing qualifies anywhere.

Which is why the lens asks whether the truth EXISTS and has an owner, not whether it’s clean. Messy is fine. Undocumented is fine. Unknowable is not.

Lens 4: Evidence. Can you measure this today, before anything changes?

Check whether the workflow is measurable right now. A unit of work volume, touch time, exception rate, cost per unit, cycle length, whatever the business already tracks.

Without a baseline you cannot prove the change. And without proof your sponsor won’t fund a second workflow, which turns this into one shot instead of a program.

Use a metric the business already tracks. Anything you invent for the occasion won’t survive being pressure-tested by finance or the board.

Invoice exception handling is a good example of a workflow that arrives pre-measured. The industry benchmark sits at an 18.4% exception rate, $9.84 cost per invoice, and 8.2 days processing time. You don’t have to build a measurement practice from scratch to know where you stand.

Lens 5: Owner. Who runs this after everyone leaves?

Name who owns this after the delivery team is gone and whoever built it has moved on. You need a process owner, a data owner, and an executive sponsor who will make the go/no-go call on the evidence.

A workflow with no owner does not qualify, however good the demo looks.

People skip this lens because it feels like paperwork. It isn’t. Security researchers surveyed 418 IT and security professionals and found 82% had AI agents in their environment they didn’t know about, with only 21% running any formal decommissioning process. They call the result “retirement debt,” agents lingering long past their usefulness while still holding live credentials. That survey was commissioned by a vendor selling the fix. The one-line version is still theirs and still right: an agent that was never formally onboarded is unlikely to be formally retired.

Every agent you deploy should have a defined lifetime, so nothing outlives its usefulness with a live connection to your systems still attached.

How to actually run the lenses

Don’t run this once on one idea. Take 3 to 5 candidate workflows and put every one through all five lenses. Then rank them on value, feasibility, risk, and what each one lets you reuse next. Then pick ONE.

That last criterion is the one people leave out, and it’s what turns a pilot into a program.

Disqualify immediately. Any candidate with no real users, no real data, no baseline, or an ask that amounts to redesigning the organization goes to the backlog rather than the shortlist.

Two categories skip the lenses, because the answer is already yes and a scoring round is wasted effort:

• Software development. That’s construction with a gate at every step.
• Work so repetitive that a script does it. Write the script.

If you genuinely can’t get the right people in a room for this, that’s the real finding and it’s worth surfacing to whoever sponsors the work. Failing that, mine what you already have: call data for themes and frustrations, product data, member data. Then ask the right people narrow questions instead of open ones.

Concentration is the part the research actually supports. BCG found leading companies focus on an average of 3.5 use cases versus 6.1 for everyone else, and anticipate 2.1x greater ROI. Fewer and deeper beats more and thinner.

What the lenses do not do

Push back on this and there’s one objection that actually lands. You’ll hear it, so here’s the real version of it.

The argument runs: the properties that make a workflow selectable are the same ones that make the result meaningless. Pick something low-risk and isolated and you’ve removed the exact pressures that determine whether anyone adopts it. Clean the data before you pilot and you’ve invalidated the test. Run it without urgency and you get polite participation from people who drift away by week three. Christian Buckley makes this case well, and he’s pro-pilot, which is what makes it worth answering.

He’s right, and the answer is to be honest about what the loop produces.

It produces a decision. Proof that this scales across the organization is a different claim and it isn’t available in four weeks at any price. Tell your sponsor you’ve proven the business case and you’ve oversold it.

There’s a related objection with real economics behind it. The productivity J-curve, from Brynjolfsson, Rock and Syverson, shows that general-purpose technologies require complementary investments that are intangible and badly measured, so early measurement systematically UNDERSTATES value. A disciplined go/no-go gate is therefore biased toward “no” exactly when the investment is the one that eventually pays.

That doesn’t kill the loop, but it should change how you read a marginal result. A clear failure is a clear failure. A near-miss on a workflow where you learned a lot about your own data is a different animal, and treating them the same will make you stop too early.

The seven-step loop

Once you have the one workflow, this is the sequence. What tools you use matters less than running the sequence. Organizations that give it real attention get value out of it. The ones that skip steps get a demo.

Step 1: Frame it

Name the workflow specifically. Not “AI for finance.” Invoice exception handling. Tier 1 support triage. Contract redlining.

Name one sponsor who owns the outcome and holds the authority to stop it or scale it. Tie it to a metric the business already tracks. And write down the non-goals, what’s explicitly out of scope, before you start. Scope creep shows up in week two and derails things.

Step 2: Diagnose it

Two things happen before anyone touches a keyboard.

Sit down with the people who run this workflow today. You’re after what never made it into the process doc: the exceptions, the judgment calls, the workarounds everybody uses and nobody wrote down.

Then agree the numbers as they stand today. Volume, cycle time, quality rate, rework, all measured over the same window. Everyone who’ll later validate the result signs off on the baseline now. Written down in advance it becomes a fixed finish line, and you’d be surprised how often people try to move one that isn’t.

Step 3: Build the evaluation harness

This is the step teams skip, and it decides whether anything you measure later means anything. In our experience it’s also where most pilots quietly die.

Before the agent exists, assemble a set of real cases with answers your own experts have agreed are correct. Then write down the specific ways this workflow can fail: a wrong number, exposed data it shouldn’t touch, false confidence about something it doesn’t know.

You can’t trust a test you design after you’ve already seen the answers.

This is stricter than normal practice in AI, and worth knowing that. The mainstream approach to evaluating language models is error-analysis-first: ship something, look at what it gets wrong, write tests for that. What I’m describing is closer to how science handles the same problem. Pre-registration in research, where you commit to your hypothesis and analysis before collecting data, exists because “preregistration prevents us from tricking ourselves.” Physics goes further with blind analysis: quarantine the region where you expect the signal, validate your method on everything else, then open the box. Some experiments have a separate team inject fake events so analysts can’t tell which results are real until the analysis is frozen. The Higgs search worked that way.

It can also be over-built. The leading practitioner on evals recommends labeling 100 to 200 examples per failure mode and budgeting 60-80% of development time on evaluation, which is not happening at a 150-person association in four weeks. Anthropic’s own guidance is more achievable: 20 to 50 simple tasks drawn from real failures is a great start, and premature sophistication is a trap. Start there.

For what it’s worth, almost nobody does this at all. In a survey of 157 organizations, only 5% said they fully trust their automated evaluation, and half had shipped an agent that passed its evals and then failed a customer.

Step 4: Pilot it

Ship the smallest version that touches real data, inside 2 to 4 weeks.

The system proposes. A named person decides. Nothing writes back to your system of record yet. It runs alongside the existing process so you have a live comparison rather than a memory, and every run leaves a trail someone can read back later.

This pattern has a name in medicine and a case worth knowing. Clinical teams call it a silent trial: the model runs on real cases in real time while clinicians stay blinded to its predictions. In one published example, a model that scored 0.90 AUROC on its held-out development data scored 0.50 in the silent trial. Sensitivity 1.00, specificity 0.0. Useless, and it looked excellent right up until it ran on live data.

The proposal-only phase was the only thing standing between that model and a clinical decision.

It isn’t an exotic ask either. AWS ships shadow testing as a product feature. Regulators already treat human oversight as a design constraint rather than a disclaimer: the EU AI Act requires that a person can override or halt a high-risk system, and explicitly names the tendency to over-rely on output. US rules on clinical decision support turn on whether the software lets a professional independently review the basis for a recommendation so they don’t rely primarily on it. Same rule, written by lawyers.

Know the failure mode. In a study of 27 radiologists reading mammograms, when the AI’s suggestion was wrong, accuracy fell from around 80% to under 20% for less experienced readers, and from 82% to 45.5% among the most experienced. A human in the loop is only a control if the loop is designed so the human can actually exercise judgment. As a design constraint it works. As a line in a policy document it does nothing.

Step 5: Measure it

Three numbers.

Throughput. How much got handled without a person touching it.

Quality. First-pass acceptance, rework, how often a person overrode the output, and anything that would have caused real harm.

Cost to run, set against value somebody has agreed is real.

That last one carries more weight than it looks. Hours saved only count as dollars once someone with budget authority has said where those dollars show up. People conclude they’ve increased employee productivity, and you’re still paying that employee exactly the same. Align on where the dollars land before you count them, or you’ll present a number the CFO dismantles in a sentence.

Step 6: Decide

Bring the sponsor and the people who run the workflow day to day into one room. Put the baseline and the pilot results side by side. Then make the call: continue, adjust, stop, or expand. On evidence rather than momentum.

If the numbers don’t clear the baseline, stopping is the right outcome. It cost you a few weeks instead of a few quarters, and it frees you to reprioritize toward a workflow that will clear it.

Willingness to stop is what makes the rest of this credible. Worth knowing that the most-quoted failure statistic in this category, the “95% of AI pilots fail” number, doesn’t say what people think it says. The underlying funnel was 60% of organizations investigating, 20% reaching a pilot, and 5% successfully implementing. Roughly a quarter of the organizations that actually ran a pilot cleared the bar. And that bar was deployment beyond pilot with measurable KPIs and ROI measured six months later, which is this loop’s own prescription.

Step 7: Scale it, then go again

Harden what worked. Monitoring, security, cost limits, and a defined lifetime for every agent.

Hand the capability to the people who run it day to day, and train them properly.

Then go back to the five lenses for the next workflow. You carry forward the evaluation approach, the access patterns, and the governance, so the second one costs less than the first and the third costs less than the second. That compounding is the point. One proven workflow is a project. A loop you can run repeatedly is a capability.

What four to five weeks actually buys you

One full loop runs four to five weeks end to end.

It does not prove business value and it does not transform the business. It gives you a decision, validated assumptions, and usually a reusable asset.

Real business impact runs on a different clock. Think 6 to 12 months. Anyone sponsoring a 1-to-3-month transformation is funding a consultant to come in, build a roadmap and a deck, and that’s about it. I’d rather say that plainly than sell you a timeline nobody hits. The best available figure, from a study of 4,000+ business leaders, puts AI deployments at under 8 months and value realization around 13 months, slightly slower than my own estimate (vendor-sponsored).

The demand on your team stays small, which is what makes it survivable alongside the day job:

• Sponsor: 30 minutes to 1 hour a week, mostly to make the go/no-go call.
• Process owner: 2 to 4 hours a week in walkthroughs and testing.
• Data owner: a few hours at the start.

No published benchmark exists for client-side hours on a pilot like this. We looked. Those numbers are what we observe on our own engagements, and nobody else seems to publish what a pilot actually costs an organization in attention.

One loop is also a stage rather than a program. Organizations move from AI curious at one end of a maturity spectrum to AI compounding at the other, and the early work is foundational: org changes, stack changes, the conditions that have to exist before sanctioned workflows make sense. MIT CISR surveyed 721 companies and found those in the first two stages of AI maturity performed BELOW their industry average on growth and profit, while the later two performed well above it. That’s correlational, and high performers may just mature faster. It points the same direction as McKinsey’s finding that out of 25 attributes tested, redesigning workflows had the single biggest effect on whether AI showed up in EBIT, and that high performers were roughly three times as likely to have redesigned a workflow rather than inserting AI into an existing one.

Where to start Monday

If your AI plan is a company-wide platform license with no clear vision, no named workflow, no baseline, and nobody owning the outcome, you don’t have an AI plan. You have a subscription.

Fixing that doesn’t take a strategy offsite. Name one workflow, specifically, where a customer or member actually feels the friction. Then write down the metric your business already tracks for it. An hour of work, no budget, and it’s the step that separates a plan from a purchase order.

Then run it through the five lenses against two or three alternatives, and if it survives, put it through the loop.

Hours saved only count as dollars once someone with budget authority has said where those dollars show up. Everything here exists to get you to that sentence with an answer.

If you want to pressure-test a candidate workflow, bring one to a 30-minute call and we’ll run it through the five lenses live. If it’s still fuzzy, bring the fuzzy version. Sorting signal from noise is most of the work.