You already use AI. So does almost every business you compete with. A Goldman Sachs survey of 1,256 small business owners found that 76% use AI and 93% of those say it has helped, yet only 14% have it fully embedded in their core operations. That gap is not a technology problem, and it is not a training problem either, whatever the survey respondents said. It is the difference between asking an AI a question and letting an AI do a job. This article explains what separates the two, what each of the three barriers owners named actually looks like once you cross it, and where to start if you want to be in the 14% by this time next year.
What Goldman Sachs Actually Measured, and Who Was Asked
The numbers come from the Goldman Sachs 10,000 Small Businesses Voices survey, fielded by Babson College and David Binder Research between 27 January and 4 February 2026 and published on 17 March. It covers 1,256 owners across all 50 US states, Washington DC and Puerto Rico. Two things about the sample matter before you read anything into it:
- It is American. Nothing in it was collected in Pakistan, the Gulf or anywhere else. Where we draw a lesson for a business in Karachi or Riyadh below, that is our reading, not Goldman's finding.
- It is not a random sample. Every respondent is a participant in Goldman's own 10,000 Small Businesses programme, which is a business-education course. These are owners who signed up to learn. If anything, the real adoption rate among businesses that did not take a course is likely lower than 76%, and the core-operations figure lower than 14%. We are speculating there; Goldman does not say.
With those caveats, the headline figures as Goldman reports them: 76% currently use AI; among users, 93% report a positive impact and 84% name efficiency and productivity as the main benefit; 67% expect AI to increase revenue; 87% say it augments rather than replaces staff; 14% say it is fully embedded in core operations; and 73% say more training and resources would help. The obstacles owners cited were concerns about data privacy and security (50%), a lack of technical expertise (49%) and difficulty choosing the right tools (48%).
Goldman Sachs 10,000 Small Businesses Voices, Jan-Feb 2026 (n=1,256, US)
Source: Goldman Sachs press release and 10,000 Small Businesses Voices insight page, 17 March 2026. Programme participants, not a random sample.
Using AI and Running On AI Are Two Different Activities
Read the 76% and the 14% together and the survey is describing one thing: most owners use AI the way they use a search engine. A person opens a chat window, asks for a draft, a summary, a translation or an idea, reads the answer, and decides what to do with it. That is real value. But nothing runs on it: if the chat window vanished tomorrow the business would be slower, and no order, invoice or customer would be missed.
Core operations is the opposite condition. It means AI is doing a job the business depends on, without a person reading every output first. From building and running systems like that, here is the test we use. AI is in your core operations only when all four of these are true:
- It runs unattended. The work happens at 3am, on a Sunday, during Eid, whether or not anyone is at a desk. A chat window only works when someone types into it.
- Its mistakes cost you money, so you measure them. A wrong summary in a chat costs a minute. A wrong decision in a workflow costs a sale, a fine or a customer. The moment an error has a price, you start counting false positives, misses and latency, which nobody does for a chat assistant.
- It has a fallback. The model provider has an outage, the API key expires, the price changes. A system in core operations has a defined answer for each of those. A chat habit has none, because it needs none.
- It has its own line in the accounts. Not a subscription buried in someone's card statement, but a cost you can put against the work it replaced and defend to yourself at year end.
That is why the gap between 76% and 14% is so wide. Getting to the first number takes an afternoon. Getting to the second takes a decision about what the business is willing to let a machine do on its own, and that decision is the hard part, not the software.
The 62-point gap between using AI and running on it is not a shortage of training. It is a shortage of jobs that an owner has decided to hand over.
The Three Barriers, Seen From the Inside
The owners in the survey named three obstacles at almost identical rates: privacy and security (50%), technical expertise (49%) and choosing tools (48%). Each one is real. Each one also changes shape once you are actually running AI in production rather than chatting with it, and the change is worth knowing about before you start.
Privacy and security (50%)
In the chat phase, the privacy worry is about what you type in. In the operations phase it is about what the system sees continuously, and the honest answer is that some jobs simply should not send data anywhere. A camera feed is the clearest case: AI Cam is built so footage is recorded and stored on the customer's own premises and never leaves the building, because for a warehouse, a clinic or a school there is no privacy policy that makes a cloud copy of every frame acceptable. The lesson generalises: before you ask a vendor how they protect your data, ask which data has to move at all. A vendor who cannot say exactly what leaves your premises has not thought about it.
The second half of this barrier is credentials. Once AI is in operations you hold API keys for it, and those are now worth stealing, as Anthropic's September threat report showed and as we covered in our guide to AI key theft. The survey's privacy worry is legitimate. It is just aimed at the wrong end of the pipe.
Technical expertise (49%)
Seventy-three percent of respondents said training would help. We would gently disagree with the diagnosis. The expertise an owner needs to put AI into operations is not technical; it is operational. It is knowing which of your processes has a clear input, a clear correct output and a tolerable cost of error. Anyone who has written a job description for a new hire already has that skill. What a course cannot teach is which of your own jobs to pick, because only you know where the money leaks.
The version of this problem that is real is building and maintaining the system yourself. For most businesses under 50 people the answer is to buy a system that already does the job and spend your expertise on defining what "done correctly" means.
Choosing tools (48%)
This barrier is getting worse, not better. New frontier models arrive most weeks, each with a benchmark table, and none of that helps an owner choose. The question that does help is boring: what specific job is this tool going to own, and how will I know within 30 days whether it is doing it? A tool that cannot be evaluated against one job in one month is a tool you are trialling, not adopting. Trialling is what the 76% are doing.
A Simple Test: Is AI in Your Operations or Just Near Them?
The same idea as a table you can run down for your own business:
| Using AI (the 76%) | Running on AI (the 14%) |
|---|---|
| Someone opens a chat window when they need help | The system starts work on its own when an event happens: a call, an invoice, a camera frame, an application |
| Output is read and edited by a person every time | Output goes straight to the customer, the ledger or the alert channel; people review a sample |
| If it is wrong, someone quietly fixes it | If it is wrong, there is a cost, so errors are counted and reported |
| Nobody knows the monthly spend to the rupee or riyal | The cost sits against the work it replaced |
| If the vendor vanished, work would slow down | If the vendor vanished, a named fallback takes over |
| Adopted in an afternoon | Adopted after a decision about what to hand over |
Two examples of the right-hand column, drawn from what our own product pages describe. In Accounts, posting a sales invoice submits it to FBR or ZATCA at the moment of issue and returns the IRN or cleared UUID and QR code onto the invoice; nobody types the invoice into a portal afterwards, and if a submission fails the error is on the invoice, not discovered at audit. In AI Cam, every camera frame is judged live, written up in plain English with a threat level, and delivered as an alert with the photo attached; a person reviews alerts, not footage. Both meet all four tests: unattended, measured, with a fallback, and with a price.
Why 14% Is the Honest Number, Not a Failing One
It would be easy to read this survey as "small businesses are behind". The evidence says otherwise. McKinsey's 2026 State of AI survey, which we analysed in our build-versus-buy piece, covers organisations of every size, more than a third of them with over $1 billion in revenue, and finds that only 6% qualify as AI high performers and only 37% report any effect on earnings at all. Large companies with dedicated AI teams are not doing markedly better at turning use into results. The 14% is not a small-business problem; it is roughly where everyone is.
The McKinsey finding that does separate winners from the rest is instructive: high performers redesign the workflow around the AI, while most others bolt AI onto the workflow they already had. That maps exactly onto the Goldman gap. Using AI is bolting a chat window onto an existing process. Running on AI is redesigning the process so the machine owns a step. There is no training course between those two; there is a decision.
Our prediction, stated as one: when Goldman runs this survey again in 2027, the use figure will be well above 76% and the core-operations figure will move far less. Adoption is nearly free and integration is not, and the gap will widen before it closes.
Where to Start: One Job, Unattended, Measured
If you want to be in the 14%, do not start with a strategy. Start with one job that meets four conditions, and give it to a system for 30 days.
- It happens many times a day. Ten times a month is a task; a hundred times a day is a process. Missed calls, invoices, camera events, job applications and WhatsApp enquiries qualify. Board reports do not.
- Correct is easy to define. "Every sales invoice reaches the tax authority with a valid IRN" is checkable. "Better customer engagement" is not. If you cannot write the pass condition in one sentence, pick a different job.
- The cost of an error is bounded. An alert sent for a cat is annoying; an alert not sent for an intruder is expensive; both are survivable and countable. Do not start with the job where a mistake ends a customer relationship.
- You can measure it in 30 days. Count what happened before and after: calls answered, invoices cleared first time, false alarms per week, hours a person spent reviewing. If the number did not move, stop and try another job. If it did, you have your first line item.
Notice what is missing from that list: the model, the vendor, the benchmark, the course. The 76% spend their time on those. The 14% spent theirs choosing the job.
What This Means If You Run a Business in Pakistan or the Gulf
The survey is American, but the gap it describes is not. If anything the region has an advantage: compliance deadlines make the choice of first job for you. In Pakistan, FBR Digital Invoicing already requires sales invoices to be transmitted at issue; in Saudi Arabia, ZATCA Phase 2 requires the same. That is a process that runs many times a day, has a one-sentence pass condition, a bounded cost of error and a 30-day measurement, which is exactly the profile above. Accounts does it from inside the ledger, and it is the most common reason a business in either country crosses from the left column to the right one for the first time.
Security is the other natural first job, because the cameras already exist and nobody watches them. AI Cam starts free on one camera with no card and no trial clock, which makes it a reasonable 30-day experiment on our own terms: does the alert count in week four look like something you would pay for? If not, you have lost a month and learnt what your cameras actually see.
For the phone, SmartLine answers in Urdu, English and Arabic and is the third job on that list. Whichever you pick, the goal is one process the business now runs on, measured, with a fallback and a cost line. That is what the 14% have. If you want help choosing which of your own jobs fits, talk to us; we would rather help you pick the right first one than sell you three.
