On 2 September 2026 the Financial Times published a survey of how AI is genuinely being used inside six industries — consulting, oil and gas, banking, law, pharmaceuticals and advertising — based on its reporters talking to people doing the work. It is one of the more useful pieces of AI reporting this year, mainly because it includes the parts that are not going well.
Most coverage of enterprise AI is useless to a small business owner. It describes systems that cost millions to build, staffed by teams larger than your entire company. Reading it as a to-do list is how people end up buying software that solves a problem they do not have.
So this is not a summary. It is a translation: what in that reporting actually transfers to a business with between two and fifty people, what does not, and — most importantly — the four failure patterns that show up identically at every size.
What the big firms are actually doing
Strip out the industry specifics and the same pattern appears six times: AI is being used to shorten the run-up to a decision, not to make the decision. Almost nothing in the reporting describes AI producing final work that goes out unreviewed. It describes AI collapsing the boring part before a human does the judgement part.
| Industry | What they're really using it for | The small-business version |
|---|---|---|
| Consulting | Searching decades of internal decks and past projects so a consultant can get up to speed on a new market in hours instead of days | You already know your business. Skip this — it's an enterprise-only problem |
| Oil & gas | Routine contract review. Shell says ~90% of its 5,000 annual NDAs no longer need a lawyer | First-pass review of supplier terms, leases and subcontractor agreements — then a human on anything that matters |
| Banking | Summarising every prior client communication so a banker can walk into a meeting with 30 minutes of prep | Pull the history on a client before a call or a site visit. Genuinely transferable |
| Law | Reading 400 agreements to find how one clause is treated — work that used to be an all-nighter | Any "find the thing buried in the paperwork" task. High value, needs verification |
| Pharmaceuticals | Cutting drug discovery timelines from 4–5 years toward 12–18 months by processing research data at new scale | Doesn't transfer. Ignore it |
| Advertising | Small firms making and distributing campaigns without hiring a large agency; adapting one asset for many audiences | The single biggest opportunity on this list for a small business |
Two of those six are worth your attention. One of them — advertising — is arguably the best small-business AI opportunity currently visible. We will come back to it. First, the part most people skip.
The four failure patterns that scale down perfectly
Every industry in the FT's reporting described a drawback. Four of them are not enterprise problems at all — they are human problems that happen to be visible at enterprise scale first, because those firms adopted earlier and have more people to observe. If you are about to spend money on AI, these are the four things most likely to eat it.
1. Token cost creep — the price you sign up at is not the price
The most under-reported line in the whole piece: consulting firms that spent 2025 pushing staff to use AI are now worried about the bill. Some smaller firms have imposed usage limits, or removed AI access from people who seem to use it less. In advertising, executives describe "token inflation" colliding with agencies that still bill clients by the hour — costs rising while the revenue model stays fixed.
This happens because serious AI pricing is usage-based. You are billed by tokens — roughly, the volume of text in and out — so the cost rises with adoption. That is the exact opposite of how a small business is used to buying software, where a seat is a seat and the invoice is the same every month.
Assume any AI tool priced per-usage will cost more in month six than month one, because the whole point is that people use it more. Set a hard monthly cap on day one, put AI in your budget as a variable line rather than a fixed one, and review real spend at 60 days. Our AI implementation cost guide breaks down what the realistic ranges look like by business size.
2. Usage theatre — when adoption becomes the target
The single most instructive detail in the article is a consultant explaining, on an industry forum, that their employer monitors AI usage and wants to see more of it. So on days they have no real reason to use it, they ask the chatbot something like "Why are we here?" — purely to register activity. In their words, they get credit for wasting tokens.
“If I haven’t had a reason to ask it a question all day, I will ask Claude something like, ‘Why are we here?’ That way, I get credit … for wasting tokens.”
— Consultant, quoted by the Financial TimesThis is what happens when a tool is mandated instead of aimed. It is not a big-company disease. The small-business version is quieter and more expensive: an owner reads enough AI content to feel behind, buys a subscription, tells the team "we should be using this," and three months later nobody can name a single number that moved. The tool gets cancelled, and the owner concludes AI does not work — when what actually failed was that it was never pointed at anything.
The fix is unglamorous. Before you pay for anything, name the specific problem it is supposed to fix and the number that will tell you whether it did. Missed calls per week. Quotes sent with no reply. No-shows per month. Hours spent on a task you can time. If you cannot name the number, you are buying usage theatre.
3. Hallucination is a liability problem, not a quality problem
Several law firms have now been censured by courts over AI-generated errors in filings. Sullivan & Cromwell — an elite firm with resources most businesses cannot imagine — apologised to a New York federal judge this year for misquoting the US bankruptcy code and citing cases incorrectly, errors caused by its use of the technology. The opposing firm, Boies Schiller Flexner, caught them; BSF had itself filed a document against Amazon the previous year containing what one of its own partners called "material citation errors."
If firms of that calibre, with review processes and professional indemnity insurance, shipped fabricated citations to a federal judge, the assumption that you will notice a subtly wrong number in an AI-drafted quote is not a safe one.
The practical rule: anything containing a number, a legal term, a date, a price or a factual claim about a third party gets a human check before it leaves the building. Anything internal, exploratory or first-draft does not. Most people apply the same level of caution to everything, which is why they either check nothing or use AI for nothing. Our guide to AI's legal cautionary tales goes through the specific ways this has gone wrong for businesses and what it cost them.
4. The productivity paradox — saved time does not stay saved
Asked directly whether AI means junior lawyers are getting more sleep, Sam Newhouse of Latham & Watkins gave the most honest answer in the piece: "I think that is probably a debatable topic." Because they can now go further and deeper on analysis, there is additional work that they can now do. If a client is making a significant investment, why wouldn't you turn over that further stone? The teams, he said, are still incredibly busy.
This is the finding that should reset expectations for anyone buying AI to "get time back." AI does not hand you hours by default. It raises the achievable standard, and the work expands to meet it. That is not necessarily bad — turning over more stones is genuinely better work — but it is not the same as going home earlier, and most small business owners are buying it expecting the second thing.
Banking's version of the same problem is subtler and worth naming: senior bankers worry that over-reliance strips out the intuitive feel for numbers that used to come from doing the laborious assembly by hand. In a small business, that intuition is the business. Be deliberate about which judgement you are willing to stop practising.
Two second-order effects from the reporting that nobody plans for. First, in advertising, teams travel at the speed of their slowest member — one holdout stalls the whole workflow, and in a team of eight that is a much bigger fraction than in a team of eight hundred. Second, in law, some clients became so worried about their data reaching AI models that they used technical controls to stop firms downloading or printing documents at all — forcing firms to add people and work less efficiently than before AI existed. Your customers may have opinions about this too.
The one finding that genuinely favours small business
Buried in the advertising section is the most commercially useful line in the article for anyone reading this site. Small companies can now use AI-powered systems that bring together the tools to make and distribute advertising without the expense of hiring a large agency — letting start-ups and independents compete against firms using AI to scale content across big central functions.
That is a real structural shift, and it is worth understanding why it favours you rather than them. Look at what the enterprise examples in the article are actually spending money on: McKinsey's Lilli, Bain's Sage, PwC's ChatPwC, JPMorgan's LLM Suite. Every one of those is a multi-million-pound project to solve a single problem — we have decades of scattered institutional knowledge and our own people cannot find it.
You do not have that problem. Your entire business — every client, every price, every process, every lesson learned — fits comfortably in a context window. The thing large firms are spending fortunes to rebuild, you already have. That is the asymmetry, and it means the sensible small-business AI strategy looks nothing like the enterprise one: fewer tools, aimed at named problems, applied to customer-facing work where the return is measurable.
The honest read of a year of real enterprise AI use is that the wins are narrower and the costs are stickier than the marketing suggests — but the wins are real, and the ones that transfer downward best are the customer-facing ones: preparation before a conversation, follow-up that would otherwise not happen, and marketing production that used to require an agency. Ignore the internal-knowledge arms race entirely. It is a problem you are lucky enough not to have.
What does not transfer — three things to actively ignore
- Bespoke internal models. Lilli, Sage, ChatPwC, LLM Suite and the like are custom builds with budgets and engineering teams behind them. The small-business equivalent is a well-organised set of documents and a good off-the-shelf assistant. Anyone selling you a "custom AI trained on your business" for five figures is selling you an enterprise solution to a problem you can solve for the cost of a subscription and an afternoon of tidying.
- Synthetic focus groups. Market researchers are now running thousands of AI "personalities" as simulated respondents to test ideas — with real decisions made off synthetic answers. Fascinating, unproven, and irrelevant at your scale. You can talk to ten actual customers this week for free, and their answers are real.
- The frontier science. Seismic decoding, drilling trajectory optimisation, biomarker identification in drug trials. These are the examples that make AI coverage exciting and they have nothing to do with running a shop, a clinic, a restaurant or a trades business. Their presence in an article is not evidence that you are behind.
A five-point sanity check before you spend anything
- Name the number. One measurable problem per tool, decided before purchase. If you cannot state what will change, do not buy it.
- Cap the variable cost. Set a hard monthly limit on anything priced by usage, and diarise a spend review at 60 days.
- Decide your verification line. Write down what always gets a human check — anything with a number, a price, a date or a claim about someone else — and what never needs one.
- Decide what you will stop doing. If you do not name the task the saved time comes out of, the time will not appear. It will turn into more work at a higher standard.
- Check your slowest link. In a small team, one person quietly not using the tool is enough to keep the old process running alongside the new one, which costs more than either alone.
If you want the fuller version of this argument with the numbers behind it, start with our honest assessment of whether AI is worth it for a small business, then what it actually costs. If you already know AI belongs in your business and the question is sequencing rather than justification, what an AI roadmap is and whether you need one covers how to order the work. And if you want the department-by-department view of where tools currently exist, see AI tools by department.
Common questions
How are businesses actually using AI at work in 2026?
Far less glamorously than the headlines suggest. Across consulting, banking, law, energy, pharmaceuticals and advertising, the dominant real-world use is compressing preparation time — searching internal knowledge, summarising long documents, producing first drafts, and reviewing routine contracts. Shell says roughly 90 per cent of its 5,000 annual non-disclosure agreements can now be handled by AI rather than a lawyer. Bankers use it to prepare for a client meeting in 30 minutes instead of hours. Almost none of the widely reported use is AI making the final decision — it is AI shortening the runway to a human one.
Does AI actually save you time?
Not automatically. The most honest finding in the field is that saved time tends to get reinvested rather than banked. Sam Newhouse of Latham & Watkins, asked whether AI means lawyers get more sleep, called it a debatable topic — teams can now go further and deeper on analysis, so the extra capacity becomes extra work rather than free hours. For a small business the lesson is that the hours only materialise if you consciously decide what you will stop doing. If you do not name the thing you are giving up, AI raises your standard instead of lowering your workload.
Why do AI costs keep going up after you sign up?
Because most serious AI pricing is usage-based, not seat-based. You are billed by tokens — roughly, the volume of text going in and out — so cost rises with adoption rather than staying flat. This is now a live enough problem that some consulting firms have imposed usage limits or pulled access from lighter users, and advertising executives describe token inflation squeezing agencies still billing clients by the hour. Budget for a variable line item, set a hard monthly cap on any usage-priced tool, and review actual spend at the 60-day mark rather than assuming the sticker price is the price.
What is the biggest mistake small businesses make with AI?
Mandating usage instead of assigning it to a problem. When AI adoption becomes a target in its own right, people generate activity to satisfy it — one consultant admitted to asking a chatbot pointless questions purely to register usage for their employer. Small teams do a quieter version of the same thing: a subscription gets bought, everyone is told to use it, and three months later nobody can name a number that moved. The fix is to attach every tool to one measurable problem before you pay for it, such as missed calls, unanswered quotes or no-shows.
Can a small business compete with a large company using AI?
In some areas, yes — and the gap is narrowing fastest in marketing and production. Small firms can now use AI-powered systems to make and distribute advertising at a level that previously required hiring a large agency, letting independents compete with far bigger central functions. The structural advantage is that most enterprise AI spending goes on solving a problem you do not have: making decades of scattered institutional knowledge searchable. A small business already knows its own business. That is a head start, not a handicap.
Questions about this analysis, or want a plan built for your specific business rather than a generic one? Get a free report →