Back to Guides
Enterprise Software 15 min read

The case for publishing your own internal AI wins

The case for publishing your own internal AI wins

Publish the win as a record, not a story. One workflow, one named author, the number before, the method that produced the number after, and a plain note on what else could explain the change. Three of those on a dated internal page will do more for next year’s AI enablement budget than any case study a vendor can send you. They are the only evidence that is about your company.

Most teams skip this and leave the win in a chat thread or an all-hands slide. Both vanish. Nine months later someone asks whether the training paid for itself, and nobody wrote anything down while the people who did the work still remembered the before.

Why the public evidence cannot fund your budget

The most quoted number in the field is one you cannot audit. MIT’s NANDA initiative reported in its 2025 GenAI Divide paper that 95% of organizations see no measurable return on their generative AI spending, and the figure went everywhere. The objections followed: the sample behind it was never published in full, and any pilot short of full production was scored a failure, which marks a deliberate feasibility test as a loss. Several of the louder critics also sell AI services, so the dispute is no more disinterested than the claim. The broad reading holds. The 95% does not travel.

Next to it sits a corpus written almost entirely by the party selling the enablement. Not a conspiracy, a sampling problem: no firm publishes the engagement that went sideways, so what is public is a list of survivors with the denominator removed.

The best-documented examples have a second defect. Microsoft’s February 2024 account of its first year with Copilot is detailed and candid, reporting about a 12% cut in case handling time across nearly 10,000 support agents in a rollout covering more than 200,000 employees. A 180-person company cannot use any of it. That one support function is larger than the whole company reading this.

Nor can you look sideways at peers, because most are not there yet. The Census Bureau’s Business Trends and Outlook Survey, covering December 2025 through May 2026, put AI use among US businesses between 17% and 20% overall, rising to 37% at firms with 250 or more employees. At the size band that signs off on training spend, the comparison set is thin and quiet.

One admissible source of evidence is left: a record you wrote. Our case for treating enablement as capability rather than license count makes that argument at the level of strategy. Here it is a filing habit.

What a published win is, and what it is not

A published win is a dated page of 500 to 700 words on one workflow, written by the person whose work changed, held where a colleague can find it without asking.

Four things it is not:

  • Not a testimonial. “This saved me so much time” is a feeling, and a finance team discounts feelings first.
  • Not a chat thread. Undated, unfindable in a month, and the number moves with each retelling.
  • Not an all-hands slide. No method section, so nothing on it can be checked later.
  • Not a dashboard panel. Usage data answers whether people opened the tool, a different question, and our guide to an adoption page a manager will open covers that surface.

The unit is a workflow, not a person and not the program. Mixing those is the most common error here: an entry that opens on one billing task and closes as a verdict on the company proves nothing about either.

The eight fields

The field list is the whole method, kept short enough that an operator fills it in one sitting.

Field What goes in it
1. Workflow and owner The specific task, its volume per month, and the named person who does it
2. Baseline The before number, how it was measured, and the date window it covers
3. What changed Which assistant, where it sits in the workflow, and what a person still checks
4. Enablement wrap Training hours received, who answered questions afterward, what was read
5. Result The after number, the date window, and the method that produced it
6. Rival explanations What else could account for the change, and why you ruled each one out or did not
7. Cost License seats, training hours, and any review time the new version added
8. Transferability Which other teams share enough of this workflow to try the same thing

Field 8 turns a write-up into more than a nice page. Without it, every team starts from zero and the company runs the same pilot nine times. With it, the second attempt begins from a known shape, and a short list of checked entries becomes the catalog that ends the pilot treadmill.

The two fields that do the real work

Fields 2 and 6 are the ones that get skipped, and they are the only two that make the entry hard to dismiss.

Field 2, the method behind the before number. A recalled estimate and a timed sample are not the same evidence. “It used to take me about three hours” is the weakest input available, and it runs generous, because the remembered version of a task includes its worst instance. A timed sample of six to ten real cases over two weeks costs almost nothing and changes the standing of everything built on it. Our breakdown of how to measure hours actually returned traces how an hours figure inflates between stopwatch and slide.

Field 6, the rival explanations. This one is borrowed, and the loan is worth naming. Case study research has a long discipline of stating what else might explain a result, and the version that fits a one-page entry is four lines:

Rival explanation What it asks Cheap check
Novelty Did the gain come from attention rather than skill? Re-measure at week 12, not week 3
Selection Was the pilot team already the strongest? Name who volunteered and who did not
Process redesign Did the workflow get cleaned up at the same time? List every change made in the window, tool or not
Measurement swap Did volume rise while quality fell? Report one quality number beside the speed number

Four honest lines beat a bigger claim, because they are the lines a CFO would raise anyway. Writing them yourself turns an objection into a credential. When a claim has to carry more weight than that, our piece on separating training effects from tool effects sets out the designs that support one.

Who writes it, and what it costs

The operator drafts it. Whoever ran the enablement program edits it. Reverse that and the entry reads as program marketing, the same defect that makes the public corpus useless.

Budget one operator hour per entry: about 40 minutes to draft, 20 to answer the editor. Total employer cost for civilian workers averaged $49.46 per hour worked in June 2026 in the Bureau of Labor Statistics release, so the sums are easy to put in front of a sponsor.

Cadence Entries per year Operator hours Editor hours Cost at $49.46
Two a month 24 24 12 about $1,780
One a month 12 12 6 about $890
One a quarter 4 4 2 about $300

Against a workshop that runs $2,000 to $15,000 at market rates, a year of records costs less than the rounding error on the training. What stops it is not money but ownership: with nobody assigned, the first entry waits for a quiet week that never comes. Name an owner and a date, or do not start.

Publish the losses or inherit the bias

A record of nothing but wins reproduces, inside your own company, the flaw that makes vendor case studies worthless. If twelve entries hold twelve successes, someone is filtering, and the next reader knows it.

The abandoned-attempt entry is shorter, four fields: the workflow, what was tried, where it broke, and what would have to be true for a retry to make sense. Half a page. People read these first, and they save the most money, because the fifth team to consider the same bad fit now stops before buying.

A record that carries failures also gets honest submissions; one known for triumphs only gets fed inflated numbers, and then it is not a record.

Where it lives, and the cadence that holds

One page per entry, one index listing every entry by workflow and date, both in the place your team already searches. If reading it means remembering a second tool, it goes unread.

Large companies build this as a product. Microsoft’s internal version pairs a scenario library of use cases and metrics with a measurement dashboard, demo videos from volunteer champions, and badges for finishing the training. At 180 people the same job is a shared folder, an index, and someone whose calendar says Friday afternoon. Either version stays alive on the distributed mechanism in our framework for reinforcement after training, so where a champion network exists, the entries are its natural output.

Monthly beats quarterly. Two entries a month keeps the habit warm and the before numbers fresh; a quarterly rhythm has every author rebuilding a baseline from memory. Watch the volume, though: a record that arrives as a weekly broadcast goes unread within a quarter, like any internal campaign.

Four jobs the record does later

Entries get written for one reason and used for four.

  1. Budget defense. A renewal talk goes differently with six dated entries and their methods on the table. They also fill the quarterly page our rethinking of the board-level AI report argues for, and hold up where a usage chart cannot.
  2. Curriculum. Field 3 of each entry is a worked example from your own documents, better teaching material than any generic exercise.
  3. Reinforcement. A named colleague’s entry about a task the reader also does is the most persuasive internal asset there is, and it costs an hour.
  4. External proof, carefully. An entry can become a public case study later, with the author’s consent, a confidentiality check, and one rule held firmly: never publish a number you could not reproduce on request. The case for keeping a record collapses the moment you put out the unfalsifiable claim you refused from a vendor.

How those feed the wider set of AI enablement metrics is the subject of our measurement guide for proving capability moved. The record is the raw material, not the model.

How to tell the record is working

Three checks, each answerable in a quarter, none of them a page count.

  • Replication. How many entries were cited as the basis for another team’s attempt? Zero after ten entries means the record is a diary, not an asset. Field 8 is usually the thing missing.
  • Challenge survival. How many entries survived a finance or sponsor question without a revised number? A low rate points at field 2 every time.
  • Loss share. What fraction of entries are abandoned attempts? Under one in six, assume filtering.

Three to five checks is the ceiling. A record graded on fifteen measures becomes its own reporting burden, the failure it exists to prevent.

A worked entry at 180 people

Written as a hypothetical, because an invented client name would make the mistake this article argues against.

Imagine a client services team of eleven at a 180-person firm. Workflow and owner: first-draft proposal writing, about 14 a month, two senior associates. Baseline: a timed sample of six proposals over two weeks in March, median first-draft time three hours ten minutes, clock stopped for interruptions. What changed: an assistant drafts sections two through five from a scoping call transcript and three prior proposals; a person writes the pricing section and reviews the rest.

Enablement wrap: one half-day session on applying an assistant to the team’s own documents, plus a shared channel for questions. Result: a timed sample of eight proposals in June gave a median of one hour fifty minutes, measured the same way, nine weeks after the session.

Rival explanations: the proposal template was also shortened in April, which likely accounts for part of the gain and is noted rather than netted out; both associates volunteered, so selection is live; win rate held steady, but the sample is too small to prove quality held with it. Cost: two seats, six hours of training time, about fifteen minutes of added review per proposal. Transferability: the renewals team drafts from the same transcripts and is the obvious second try.

Nothing in that entry is impressive. It is checkable, which is the point. Six of them across a year make a case for a second phase that rests on something other than goodwill.

Frequently asked questions

What is the AI enablement case study method?

As practitioners use the term, it is a loop: baseline a workflow before any tool arrives, run a defined attempt with a stated success threshold, then write a short record of what happened, what it cost, and which teams could copy it. The academic sense of the phrase means case study research design, a far heavier discipline.

Who should write an internal AI win?

The person whose work changed. The enablement owner edits for structure and presses on the method, but an entry written start to finish by the program’s sponsor reads as promotion and gets discounted as such.

What if the result was negative?

Publish it as an abandoned attempt: workflow, what was tried, where it broke, and the condition under which a retry makes sense. These get read more than the wins and prevent more waste.

Do we need a control group for this to count?

No. A stated method, a dated before and after, and four honest lines on rival explanations will carry an internal decision. Control groups and staggered rollouts matter when a claim has to survive outside scrutiny or back a large spend.

How many entries before the record is useful?

Three make a budget conversation possible. Six to eight across a year start to show pattern, which is when field 8 pays and teams stop repeating each other’s pilots.

Can we publish an internal win externally?

Yes, with the author’s consent, a confidentiality check, and the discipline of publishing only numbers you could reproduce on request. Restating an internal estimate as a public statistic is what made the existing corpus unusable.

What does the record cost to maintain?

About one operator hour and half an editor hour per entry. At the June 2026 average employer cost of $49.46 per hour worked, a monthly cadence runs under $900 a year, small against the $2,000 to $15,000 market range for the training it documents.

Should entries include dollar figures?

Only where the conversion is defensible. Hours and quality measures hold up under challenge; an hours-to-dollars conversion rests on an assumption about what the freed time bought, and that assumption belongs in the entry, not hidden inside the number.

Key takeaways

  • The public evidence cannot fund your program: its headline failure rate is contested, its success stories are written by sellers, and its largest examples sit at a headcount no mid-sized company can map onto.
  • A published win is a dated 500 to 700 word record of one workflow, written by the operator. Fields 2 and 6, the measured baseline and the other explanations, are the two that get skipped and the only two that make it hard to dismiss.
  • Publishing abandoned attempts is not optional. A record of pure wins recreates the survivorship problem that makes vendor material worthless.
  • At one operator hour per entry, a year of records costs less than the rounding error on a single workshop, so the constraint is ownership and a date, not budget.
  • Grade the record on replication, challenge survival, and the share of entries that are losses. Three checks, not fifteen.

If you are being asked to show what your AI licenses bought and have nothing written down yet, book a free AI-readiness call. It is a direct way to settle which workflow to baseline first and what the first entry should hold.

Last Updated: Nov 14, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles