Your Pilot Worked. Now What? The Scale-or-Stop Decision for Your AI Rollout

The pilot worked. The business question starts now

Picture a ten-person business in Bengaluru — a home-services firm, say, or a small coaching institute. For three weeks, one team member has been using an AI assistant to answer routine enquiries from the business's own price list and service notes. The answers come back in seconds, the tone is right, and the owner has stopped dreading the enquiry inbox. The demo, by any honest measure, worked.

This is the moment most small businesses get wrong — not because the pilot failed, but because it succeeded. A successful pilot answers one question: can the task be done this way? The rollout asks a different question entirely: can the business run it this way, every day, at full volume, when the person who ran the pilot is on leave and the awkward enquiries arrive? Those are two different decisions, and the second one is where the real money goes. This post is about making the second decision on purpose, on one page, before the spending starts.

The distinction is not academic. It is the single clearest finding in the October 2026 research on AI adoption — and it is the reason this post exists.

What October's numbers actually say

In October 2026, coverage of consultancy BearingPoint's study travelled fast: the study itself rests on an August survey of 1,050 senior leaders across 13 countries, reported by Reuters on October 1. The headline said most companies are seeing returns from AI, yet most cannot get past the pilot. Read carefully, the study says something more precise, and more useful, than the headline.

• 74% of organisations with implemented AI reported measurable top- or bottom-line impact, according to TechTarget's coverage of the study. The Reuters report phrases the same figure as companies surveyed; the scoping used here is TechTarget's implemented-AI one.
• Fewer than one-third of surveyed companies had moved beyond pilot projects, and only 13% of companies were on track with their AI initiatives, per the Thomson Reuters report (syndicated via Q106FM, not a Reuters primary URL). TechTarget phrases the same 13% finding as scaled completely in line with the original business case.
• The two most-cited barriers were legal/regulatory barriers, named by about 40%, and integrating with existing IT systems, named by 34%.
• Leaders — the most mature organisations in the study — scaled fully as planned 47% of the time, versus 6% for Implementers, per Startup Fortune's corroboration.
• Around 24% reported AI-driven cost savings of at least 10%, versus only 4% reporting revenue growth of the same scale.

Two honest qualifications ride along with these figures. First, this is a consultancy study reported by Reuters and TechTarget, not hands-on verification — read the figures as reported, not as tested by us. Second, the study covers large companies, not ten-person businesses. The percentages will not be yours. What transfers is the shape of the finding: returns without scale, and a gap between proving value and scaling it that has little to do with whether the technology works. BearingPoint's expert summarised the gap, in Reuters' account, as proving value and scaling it being two different things.

Two famous numbers, and why both are true

If you have followed the AI debate at all, you have met two numbers that seem to contradict each other. One says most companies see a return from AI within a year. The other says almost no AI pilot delivers measurable profit-and-loss impact. Both numbers are real, both come from serious work, and the reason they coexist is the whole argument of this post.

The Google finding is corroborated by Virtualization Review's independent coverage of the same study (3,466 leaders, 24 countries, 74% of the surveyed enterprises — all with existing generative AI deployments — and 88% among agentic early adopters). The MIT finding comes from CloudZero's benchmark table, which also carries the $30–40 billion enterprise investment figure.

The reconciliation, from CloudZero's updated October 2026 benchmarks, is that the two studies measure different populations with different definitions. Google surveyed enterprises that already had AI deployed and asked whether at least one use case shows ROI. MIT required measurable P&L impact across the projects it analysed. The variable that separates the two groups is measurement capability: organisations that can attribute spend to a use case and trace it to a financial outcome show up in the 74%; organisations that cannot show up in the 95%, sometimes with good AI nobody can prove is good. Per CloudZero's account of the report, the winners combined narrow scope, deep integration, and a measurable outcome.

Both numbers come with caveats, and neither is hidden here in favour of the flattering one. Google Cloud has an interest in AI success stories, and its ROI definition is broad — revenue gains plus organisational benefits such as productivity. MIT's report describes itself as preliminary findings and has drawn methodology criticism, including an alternative funnel reading of the report's own chart (60% investigated, 20% piloted, 5% deployed), published in the aipersonalization methodology walk-through — an alternative reading to keep separate from the project-level figure, not a second finding. One more caution: secondary summaries disagree on the study's interview and survey counts, so none are cited here — only the 300+ projects figure, on which the summaries agree.

The lesson for a small business owner is blunt and practical: neither headline number is a verdict on your business. Your business has exactly one number that matters — your own, measured on your own pilot. The rest of this post builds the page that produces it.

Why small businesses feel this gap first

If the scale gap were only a large-company problem, this post would not be for you. But the small-business evidence points the same way, from the other side: adoption is broad, guidance is missing, and the missing guidance is specifically about cost and return.

• Meta's YouGov survey of 3,983 small business owners at businesses with 1 to 49 employees across nine Asia-Pacific markets including India (fieldwork August 6 to September 2, 2026; published October 14, 2026) found 84% use AI, 76% report a positive productivity impact, and 71% a positive revenue or sales impact over the past 12 months, per Meta's own release and MarTech Edition's independent coverage.
• The same survey found 86% say small businesses need more guidance to gain the full benefits, cost is the leading barrier to training at 51%, and owners asked for clearer information on cost and return on investment, and for guidance on data privacy and security.
• The everyday uses are practical, not exotic: among owners whose business used AI in the past year, 51% created or edited content, 43% ran local advertising and marketing, 38% used translation, and 37% used customer support and communications.

The same caution applies here. Meta commissioned the study and sells advertising tools to the surveyed population; Meta states the results are not nationally representative; and no India-only breakdown is published — so read the figures as APAC owners including India, not as Indian owners. The finding is directional, not a census. But direction is enough for the decision at hand: the businesses that have adopted AI are now asking, in large numbers, for help knowing what it costs and what it returns. That is precisely the question the Scale-or-Stop Sheet answers — and it is the question this site's AI Strategy & Readiness Consulting work exists to answer before the spending starts.

What changes when you scale: the cost lines a pilot never carried

A pilot is cheap in ways that have nothing to do with price. It runs at low, controlled volume. Someone watches every output. The awkward cases get handled by hand and nobody counts them. The person who built it is in the room. Scale removes every one of those cushions at once — and the new cost lines arrive as categories, not as a single bill.

• Volume: real traffic, with spikes and retries, replaces the pilot's handful of test cases. Per-unit costs can fall while total spend rises, because consumption volume grows faster than unit prices fall — the mechanism documented by YodaPlus.
• Review: the pilot's informal glance at every output becomes a staffed review queue for the answers the AI gets wrong. Handling wrong answers — spotting them, correcting them, keeping them from reaching customers — is a cost line of its own, not an afterthought.
• Upkeep: the pilot's single prompt becomes a maintained asset, and the business starts paying for the checking around it — observability, evaluation, drift-monitoring, prompt upkeep, and support coverage that the pilot budget never carried, per EasyFlow's pilot-vs-production cost table (categories only).
• The budget evidence: the FinOps Foundation's 2026 State of FinOps found 73% of companies exceeded their original AI cost plans, corroborated independently by Innobu and Analytics Insight.

Two cautions before you use these categories. First, they are decision inputs only — cost categories and ownership rows on the sheet, not a guide to running observability, evaluation, or drift-monitoring. Second, watch out for a lookalike number: a separate Harness survey of roughly 700 enterprise IT leaders also reports 73%, but of AI projects exceeding budgets — a different denominator from the FinOps Foundation's companies exceeding original AI cost plans. Keep the two apart.

The discipline that ties these categories together comes from CloudZero's ROI benchmarks: time savings only become ROI when they convert to shipped work, deflected headcount, or revenue. A productivity gain nobody can trace to the P&L is not a return. That is why the sheet below starts with a baseline and a countable unit — cost per job handled, per enquiry received, per document processed — agreed before the rollout, not negotiated after it.

Build your Scale-or-Stop Sheet

Here is the one artifact of this post: a single page that turns "the pilot worked" into a decision. It has six rows. A pilot that cannot fill all six is not ready to scale, however good the demo looked. That rule is the post's whole prescription, and the steps below build the page in one sitting.

Step 1: Open a blank Sheet and copy the template

1. Open a new Google Sheet and name it "Scale-or-Stop Sheet — [the job the pilot does]".
2. Copy the template below into the first cells, one row per line. Column A is the row name, column B is your answer, column C is where the number came from (a count, a bill, a timed week — not a guess).
3. Expected result: six empty rows with those three columns, ready to fill.

The template to copy — paste it as-is, then replace the bracketed text with your own answers:

Row | Your answer | Source of the number
Baseline | [What the job costs in time or money today, per week] | [Timed week / payroll / bill]
Volume | [How many times the job actually happens per week at full scale] | [Enquiry log / order book / inbox count]
Operating cost | [Cost per attempted job at full volume, plus fix time for wrong jobs] | [Pilot bills + wrong-job count]
Owner | [One named person who runs it and fixes it] | [Name]
Exception path | [What happens to the jobs the AI gets wrong, and who handles them] | [Named handler + process]
Stop rule | [The number or date at which you pause or kill the rollout] | [Threshold + date]

Step 2: Fill the baseline and volume from your own records

1. Baseline: pick one normal week and time how long the job takes without AI — or note what it costs — per week. Use the week before the pilot started if you have those records; otherwise time a current week with the AI switched off. Write the figure and its source in columns B and C.
2. Volume: count how many times the job actually happens in that same week — enquiries answered, invoices processed, follow-ups sent. This is the full-scale number, not the pilot's number.
3. Expected result: rows 1 and 2 filled with counted figures, each with a source. If either row is a guess, mark it "GUESS" and treat the sheet as incomplete.

Step 3: Price the scaled version honestly

1. Operating cost: take the pilot's bills and divide by the number of jobs the pilot attempted — every job, right or wrong — to get a cost per attempted job. The pilot's bills are spread across all attempts; the human time to fix wrong jobs sits on top of the bills, not inside them, so nothing is counted twice. Multiply the cost per attempted job by the full-scale weekly volume from step 2. Then add the review time: count the pilot's wrong jobs per week, time how long each takes to fix by hand, scale that fix time to full volume, and add it.
2. Write the result in row 3 with its working shown in column C — pilot cost per attempted job, full-scale volume, wrong-job count, fix time per wrong job.
3. Expected result: a row-3 figure built from your pilot's own bills and wrong-job counts, not from any vendor's numbers. This post deliberately carries no vendor cost figures for you to copy; the sheet uses your numbers only.

Step 4: Name the owner and the exception path

1. Owner: write one person's name in row 4. Not a team, not "we" — one person who runs the workflow and fixes it when it breaks, including while the pilot's champion is away.
2. Exception path: write, in row 5, exactly what happens to a job the AI gets wrong — who sees it, how fast, and what they do. If the answer is "someone will notice," the row is blank.
3. Expected result: rows 4 and 5 filled with a name and a process. A rollout with no named owner is a pilot with a bigger audience, not a system.

Step 5: Set the stop rule and the decision date

1. Stop rule: pick one number or one date in row 6 — for example, a maximum cost per attempted job, a maximum wrong-job rate, or a review date four weeks out — beyond which you pause or kill the rollout.
2. Decision date: write the date, within a week of filling the sheet, on which you will sit down with the filled sheet and decide: scale, or stop.
3. Expected result: a dated decision, not an open-ended drift. The stop rule is what makes the sheet a decision instead of a diary.

Step 6: Draft the sheet with an assistant, then check it by hand

1. Paste the starter prompt below into the AI assistant you already use, fill in the bracketed facts from your pilot, and ask it to draft the six rows.
2. Read every row it produces against your own records. Correct any number it invented, any row it left as a platitude, and any owner it left unnamed.
3. Expected result: a filled sheet in your own numbers, checked by you, with a decision date within a week. The assistant drafts; you decide.

The starter prompt — copy it exactly, fill in the brackets, and paste it into your assistant:

I ran a small AI pilot in my business. The pilot job is [the job, e.g. answering routine enquiries from our price list]. In the pilot week it attempted [number] jobs, cost [your pilot bills], and got [number] jobs wrong, each taking [minutes] to fix by hand. At full scale the job happens [number] times per week. Today, without AI, the job takes [hours] per week.

Draft a one-page Scale-or-Stop Sheet with exactly these six rows: baseline, volume, operating cost, owner, exception path, stop rule. For operating cost, show your working: pilot cost per attempted job (bills spread across all jobs attempted), multiplied by full-scale weekly volume, plus wrong-job fix time at full volume. Leave the owner row blank for me to fill with a name. End with one sentence: the single number on this sheet that tells me most clearly whether to keep or stop the rollout. Do not invent any number I did not give you; mark anything you cannot compute from my figures as UNKNOWN.

One worked example, so the shape is concrete — and explicitly illustrative, not a reported result. A fictional Bengaluru boutique runs an enquiry-answering pilot. (Illustrative example only — every number here is invented to show the arithmetic, and no part of it is a measured result from any real business.) In the pilot week it attempts 40 enquiries: 36 answered correctly, 4 wrong. Its records work out to 10 cost-units per attempted job, and each wrong enquiry takes 15 minutes to fix by hand. At full scale the job happens 150 times a week. The sheet's arithmetic is: 10 cost-units × 150 attempted jobs = 1,500 cost-units of operating cost, plus review time — the pilot's wrong-job rate is 4 in 40, so at 150 jobs that is about 15 wrong enquiries a week, × 15 minutes = 225 minutes of fix time. The bills are spread across all 150 attempts; the 225 minutes of human fix time sits on top, so nothing is counted twice. If the owner cannot fill the exception row from the pilot's own wrong answers, the sheet is incomplete — and that incompleteness is the finding.

The rule: no full row, no rollout

The rule is short enough to pin above a desk: a pilot that cannot fill all six rows is not ready to scale, however good the demo looked. The rule earns its bluntness from the failure rates behind it.

• Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value, per Startup Fortune's handoff analysis and CloudZero's enterprise ROI benchmarks.
• S&P Global Market Intelligence found in its March 2025 survey that 42% of companies had abandoned most of their AI initiatives, up from 17% the year before, scrapping 46% of proofs of concept on average, per CloudZero's enterprise benchmarks, The Operations Bridge, and CloudServus.
• The handoff discipline proposed for this gap — a success metric, a named decision owner, and a hard date — originated as a pre-pilot contracting practice (agree the conversion terms before the pilot starts). This post adapts it into a post-pilot decision sheet, and presents it as an adaptation, not as originally post-pilot advice.

One limit on everything above, stated plainly: the Gartner and S&P figures show how often initiatives fail; they do not prove this sheet prevents failure. The sheet is a decision discipline, not a guarantee. What it guarantees is narrower and still valuable: that the decision to scale will be made from counted numbers, a named owner, and a dated stop rule — instead of from the glow of a good demo. If the filled sheet says stop, stopping is the sheet working, not the pilot failing. A stopped rollout that cost one pilot's effort is a bargain next to a scaled rollout nobody owns.

That is the work this site's AI Strategy & Readiness Consulting does before the spending starts: figuring out where AI actually pays off, in a specific business, with its own numbers — and saying stop when the numbers say stop. If your pilot worked and you are staring at the rollout decision, fill the sheet first. The blank rows will tell you what to do next.