Search the blog
Roughly a third of UK businesses with ten or more staff now say they use artificial intelligence, up from about 12% in late 2023 (ONS, 2026). In our line of work that shows up as a new kind of support call: the script an assistant wrote that deleted the wrong folder, the firewall rule pasted from a chatbot that opened a port to the world, the "quick fix" nobody can explain a month later because nobody wrote it down.
This is not an argument against the tools; the best studies show real gains. It is an argument that the habits IT people learned the hard way — back it up first, least access needed, one change at a time, write it down — are exactly what stops a fast tool becoming a fast disaster. Here is the evidence on both sides, then the ten habits and how to fit them around AI.
What the evidence actually says
The gains are genuine. In a randomised experiment, college-educated professionals given ChatGPT for writing tasks finished 40% faster and their work was graded 18% higher; access "allowed nearly everyone in the treated group to perform as well as the top humans in the control group" (Noy and Zhang, 2023). In a study of 5,172 customer-support agents, an AI assistant lifted productivity by 15% on average, with the biggest gains for the least experienced staff and for rare problems (Brynjolfsson, Li and Raymond, 2024).
The failure modes are just as well documented. At Stanford, developers with access to a code assistant "wrote significantly less secure code than those without access" — and were more likely to believe their code was secure (Perry et al., 2023). When researchers timed 16 experienced open-source developers across 246 real tasks on their own mature codebases, allowing AI made them 19% slower, even though the developers had predicted a 24% speed-up and still believed afterwards that they had been 20% faster (Becker et al., 2025). Asked verifiable questions about US federal court cases, GPT-4 produced false answers 58% of the time and Llama 2 88% of the time (Dahl et al., 2024). The NCSC puts it bluntly: these systems "get things wrong and present incorrect statements as facts" and are "often gullible when responding to leading questions" (NCSC, 2024).
The pattern is consistent. AI helps most on routine work, for less experienced people, where the output is easy to check. It helps least — and can quietly harm — on complex systems the tool does not fully understand, judged by people who trust it too much. The overconfidence is the finding. That is why the old disciplines matter.
The ten habits
1. Back up before you change anything
An assistant will happily generate a migration script, a registry edit or a bulk rename and run it in one go. A tested, known-good copy before any change turns a wrong guess into a ten-minute rollback instead of a data-recovery job. Combine: ask the tool to write the backup and restore steps first, then the change. Our 3-2-1 rule applies just as much to servers as to family photos.
2. Least privilege — for people and for the tool
Give an AI agent your administrator credentials and it can do everything you can, including the mistakes. The NCSC warns of prompt injection, where crafted input makes a model "behave in an unintended way", including revealing confidential information (NCSC, 2024). Combine: run assistants under a limited account, read-only by default, and treat their output as untrusted input, like anything else that arrives from outside.
3. Change one thing at a time
A tool that offers to "fix everything" is offering to hide the cause when something breaks. Change control exists so that when a system stops working, the list of suspects has one name on it. Combine: accept generated changes in small, reviewable pieces, each with a reason and a way back.
4. Understand the system before you edit it
The developers who got slower in the METR study were working on large codebases they already knew intimately; the tool did not know them, and reviewing its suggestions cost more than it saved (Becker et al., 2025). Combine: use AI to explain unfamiliar systems, logs and error messages — that is where it shines — and keep the decision about what to change with the person who understands the consequences.
5. Test it — and do not trust the tool's own test
In the Stanford study, the people who produced the most secure code were those who "trusted the AI less" and engaged more with their prompts (Perry et al., 2023). A generated function plus a generated test that passes proves only that the two agree with each other. Combine: write the test case yourself, or at least the expected result, before asking for the code.
6. Write it down
Half of businesses in the ONS survey report no change in headcount from AI (ONS, 2026), which means the same people are now maintaining more systems, faster. Undocumented automation is a liability that grows quietly. Combine: make the assistant produce the change note, the runbook and the rollback plan as part of every task; that is the one piece of typing it genuinely saves.
7. Measure before, measure after
The METR developers were confident they had been faster; the clock said otherwise (Becker et al., 2025). Confidence is not a metric. Combine: baseline ticket times, error rates and rework before rolling out an assistant, then look again after a quarter. If the numbers moved the right way, expand; if not, you have learned something for the price of a spreadsheet.
8. Verify anything that claims to be a fact
Version numbers, command-line switches, licence terms, compliance statements: the legal-hallucination figures above are a warning about every field where an authoritative-sounding answer is easy to produce. Combine: ask for the source, open it, read it. No source, no fact.
9. Keep a human who can say no
The NCSC's secure AI development guidelines run across four phases — design, development, deployment, and operation and maintenance — with logging, monitoring and update management named explicitly in the last (NCSC, 2023). Nothing in that framework removes the person who watches the dashboard. Combine: automate the doing, not the deciding. Approvals for anything that touches money, identity or backups stay with a named person, as in our invoice-fraud rule.
10. Know how to do it by hand
The customer-support study found the largest gains among staff with the least experience (Brynjolfsson, Li and Raymond, 2024). That is wonderful on day one and a risk on day 1,000, when the tool is offline, wrong, or quietly retired and nobody remembers how the job was done. Training existing staff is already the most common way UK businesses add AI capability (ONS, 2026) — make sure that training covers the fundamentals, not just the prompts.
| Habit | Why it still matters | How to combine with AI |
|---|---|---|
| Back up first | Wrong guesses become rollbacks, not recoveries | Generate backup and restore steps before the change |
| Least privilege | Prompt injection and leaked secrets | Limited accounts, read-only by default |
| One change at a time | Keeps the list of suspects short | Small, reviewable, reversible pieces |
| Understand the system | Experts got 19% slower on unfamiliar-to-the-tool code | Use AI to explain; humans decide |
| Test independently | Assistant users overrated their code's security | Write the expected result yourself |
| Document | Same headcount, more systems | Make the tool write the runbook |
| Measure | Felt 20% faster, was 19% slower | Baseline, then compare |
| Verify facts | 58–88% false answers on legal questions | Open the source |
| Human sign-off | NCSC: monitor and log in operation | Automate doing, not deciding |
| Manual skills | Gains concentrated in novices | Train fundamentals, not just prompts |
The strongest counter-argument
The fairest objection: most small businesses never followed the old disciplines anyway. Suppose a 12-person accountancy firm has no change control, one overworked person and a shared password. If an assistant gets that firm from nothing to something — a documented script, a backup it remembered to mention, an error explained at 6 pm on a Friday — the tool has raised the floor, and the studies showing novices gaining most support exactly that. On that view, insisting on full discipline mostly benefits people who already had the skills.
There is real force in this, and it is why we do not tell customers to avoid the tools. But the same evidence shows novice gains arrive with the largest overconfidence, and a business with no fallback skills is the one least able to notice when the tool is wrong. Raising the floor is good; removing it is not.
What this means for you
If you run a small business: let staff use assistants for drafting, explaining and summarising, and put a short written rule in place for anything that changes a system, moves money or touches customer data — a backup, a second pair of eyes, a note of what was done. At home, the same applies to any "run this command" advice, whoever or whatever gave it.
If you would like a plain-English review of where AI tools are already in use in your business and which of the ten habits are covered, that is a fixed piece of work with the findings in writing. See our business IT support overview or get in touch.
Sources
- Becker, J., Rush, N., Barnes, E. and Rein, D. (2025) Measuring the impact of early-2025 AI on experienced open-source developer productivity. arXiv:2507.09089. Available at: https://arxiv.org/abs/2507.09089 (accessed 25 August 2026).
- Brynjolfsson, E., Li, D. and Raymond, L. (2024) Generative AI at work. arXiv:2304.11771, v2. Available at: https://arxiv.org/abs/2304.11771 (accessed 25 August 2026).
- Dahl, M., Magesh, V., Suzgun, M. and Ho, D.E. (2024) Large legal fictions: profiling legal hallucinations in large language models. arXiv:2401.01301. Available at: https://arxiv.org/abs/2401.01301 (accessed 25 August 2026).
- National Cyber Security Centre (2023) Guidelines for secure AI system development. Available at: https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development (accessed 25 August 2026).
- National Cyber Security Centre (2024) AI and cyber security: what you need to know. Available at: https://www.ncsc.gov.uk/guidance/ai-and-cyber-security-what-you-need-to-know (accessed 25 August 2026).
- Noy, S. and Zhang, W. (2023) 'The productivity effects of generative artificial intelligence', VoxEU, CEPR, 7 June. Available at: https://cepr.org/voxeu/columns/productivity-effects-generative-artificial-intelligence (accessed 25 August 2026).
- Office for National Statistics (2026) Artificial intelligence in UK businesses: 2023 to 2026. Available at: https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026 (accessed 25 August 2026).
- Perry, N., Srivastava, M., Kumar, D. and Boneh, D. (2023) Do users write more insecure code with AI assistants? arXiv:2211.03622, v3. Available at: https://arxiv.org/abs/2211.03622 (accessed 25 August 2026).