01 — What happened
On 14 October 2026 the Federal Reserve published the October Beige Book, prepared at the Philadelphia Fed from information collected on or before 5 October. Its labour markets summary carries one sentence worth reading slowly:
“Five Districts reported that some firms had either reduced staffing levels (through attrition), delayed hiring, or anticipated less job creation after increasing productivity by investing in AI.”
Take the three outcomes separately. Staffing reduced through attrition. Hiring delayed. Less job creation anticipated. Not one of them is a redundancy. The same report describes little or no net change in overall employment levels across most Districts.
The District detail says the same in different words. Minneapolis IT staffing contacts reported that AI “was not causing layoffs but was dampening the creation of new jobs”. A Chicago business services contact said AI productivity gains “allowed their company to reduce employment through attrition”. A Dallas accounting contact noted that some large firms are delaying new employment contracts “because they may not need as many entry-level employees as initially anticipated, partly due to efficiency gains from AI”.
Six days earlier, on 8 October, the New York Fed published an analysis of the AI module of its 2025 Small Business Credit Survey, covering 5,248 employer firms. Some 46% said the business or its employees were using AI tools, and a further 15% planned to adopt within twelve months. Among users, 63% called AI somewhat or very important to production and 51% had integrated it partially or fully into business processes.
Then the two figures that matter most: 77% reported no change in their firm’s current labour costs from AI use, and only 31% reported increased sales from it.
02 — Why it matters
Side by side, these two publications stop contradicting each other.
Nearly half of small firms use AI. Half of those have put it inside a process. Most cannot see it in their labour costs, and roughly two-thirds cannot see it in their sales. Meanwhile the effect the Fed’s contacts actually report is expressed almost entirely in the future tense: a post not refilled, a hire deferred, a headcount plan trimmed before it was ever acted on.
That is the finding underneath both reports. AI’s return is currently arriving as cost avoided, not cost removed.
The distinction decides whether anyone ever sees the money. A cost removed appears in this year’s accounts: a line that was there in March is absent in September. A cost avoided appears nowhere. There is no accounting entry for the salary you never started paying. It surfaces, if at all, as a margin that held while volumes rose, which almost nobody attributes correctly.
That explains the survey result without reaching for hype or lag. Firms that were about to hire and then didn’t have banked something real. Firms that were never going to hire have banked nothing, and report exactly what 77% of them reported. Which changes the first question to ask before buying. Not “is the tool any good”. Were you about to spend?
03 — The economics
A time saving is not money. It becomes money only if it has an exit.
There are three exits, and a business needs one already open:
- A hire you were about to make, and now don’t. The saving is the loaded cost of that role, in full, from the month the salary would have started.
- A queue that converts. Capacity redirected into revenue-producing work: enquiries answered faster, quotes out same-day, lapsed customers actually called. The saving is the margin on work that would not otherwise have happened.
- A cost line you can close. A contractor not renewed, an outsourced function brought back in, overtime stopped.
If none is open, the saving stays inside somebody’s working day and is spent on being less busy. Not worthless — but not a return, and it will not appear anywhere you can point to. That is a prior condition to the question of how an automation should be valued at all.
BCG’s 2026 AI at Work survey, covering 11,749 workers across 14 markets, describes the failure mode plainly. Among frontline employees who use AI regularly, 42% report saving roughly a full working day a week. But 66% receive limited or no guidance on what to do with that time, and more than half do not redirect it into higher-value work. The hours are real. The exit is missing.
Take an illustrative example. A twenty-person service business handles, say, 400 enquiries a month with two coordinators on intake and admin. Suppose AI tooling removes six hours a week from each — a conservative reading of the BCG figure. That is around 0.3 of a role.
In the first case, the business was about to hire a third coordinator because volumes were climbing. The 0.3 does not cancel that hire, but it pushes it out two or three quarters. The saving is most of a salary, deferred — worth more than anything else on the benefits list.
In the second case, volumes are flat and no hire was planned. Those twelve hours must now find an exit. Put them against enquiries currently answered in two days, improve the answer rate, and the gain is commercial and measurable. Leave the exit undefined and the hours dissolve into a slightly easier week — and twelve months later the owner reports no change in labour costs and none in sales, having done nothing wrong except skip that step.
An NBER working paper from March 2026, drawing on nearly 750 corporate executives, describes a “productivity paradox, in which perceived productivity gains are larger than measured productivity gains”, attributing it largely to revenue realisations lagging. Lag is one half. The missing exit is the other, and it is the half a business controls.
04 — What businesses should do
Name the hire you are deferring, in writing, before you buy anything. A role, a month, a loaded cost. If you cannot produce all three you are not in the first case, and the project should stop being described internally as a saving.
If you are in the second case, define the exit before the build. Pick one queue that converts and currently runs behind: unanswered enquiries, quotes outstanding over 48 hours, renewals nobody has rung. State the number that should move, and in which direction. “Response under four hours on 90% of enquiries” is an exit. “Free up the team” is not.
Scale the decision to volume honestly. If your business handles 500 inbound enquiries a month manually and loses some to slow response, AI-assisted triage and drafting is worth evaluating now: the exit is already identified, measurable and revenue-facing. If it handles forty, it is not — however well the tool demonstrates.
Check whether anyone could tell either way. Ask what last month’s enquiry volume, response time and conversion rate were. If nobody can answer, no AI investment here will ever be provable — and unprovable spending is the first cut when budgets tighten. That gap comes first, and it sits behind the real cost of disconnected business systems.
Expect to feel this in capacity per person, not in a smaller payroll. The New York Fed’s data points firmly that way: AI-using small firms expect higher employment than non-users — a 33-point net expectation against 15, narrowing to 10 points once firm, owner and prior-hiring controls are applied. These are not businesses shrinking, but businesses expecting to grow without the headcount that growth used to require, which is also where the junior rung quietly gets removed.
05 — Where technology fits
For a business in the first case — about to hire, volumes climbing — the useful build is narrow and unglamorous. AI applied to one high-volume, judgement-light stage of a process that is already defined: triage, first-draft replies, pulling data out of documents that arrive in the same shape every week. The economics work because the comparison is a salary.
For a business in the second case, the first problem is often not AI at all. There is no queue to put capacity into, because the work is not visible anywhere: nobody can say which enquiries are waiting, who owns them, or how long they have been sitting. That is a system-of-record problem — a CRM, or a connected set of systems — and fixing it is what makes a later AI investment measurable rather than hopeful. Build the counter before the engine.
And the honest no. If your volumes are flat, your team is not the constraint and nothing is queueing, you probably do not need an AI system for this yet. Saved hours with nowhere to go are a comfort, not a return, and buying them is how a business ends up inside that 77%.
06 — RevenueStack’s take
Our view: the most useful line published this month is the Beige Book’s three verbs — reduced through attrition, delayed hiring, anticipated less job creation. Every one describes a cost never incurred rather than one taken out. That is what AI currently is, economically. Not a cost-reduction technology, but an option on growth: it lets a business take on more without taking on more people.
An option is only worth something to whoever can exercise it. A business with rising volumes, a visible backlog and a hire pencilled in holds a valuable one. A business with flat volumes and an underused team holds a worthless one — and that 77% figure is largely the sound of firms finding that out a year after buying.
So “what should we automate?” is the second question. The first is duller and decides everything: what were we about to spend, and on what?
Sources
- The Beige Book: Summary of Commentary on Current Economic Conditions by Federal Reserve District, October 2026 — Board of Governors of the Federal Reserve System, 2026-10-14
- AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners — Liberty Street Economics, Federal Reserve Bank of New York, 2026-10-08
- AI Is Reshaping Jobs Faster Than Companies Are Reshaping Work — Boston Consulting Group, 2026-06-03
- Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives — National Bureau of Economic Research, 2026-03-01
- AI at Work: Why Strategy Matters More Than Tools — Boston Consulting Group, 2026-06-03
