A senior associate at a recruitment firm I work with told me last month that ChatGPT had stopped being useful. Her team had tried it for screening CVs, drafting candidate emails, and writing job descriptions, and the output was generic and unimpressive. I asked to see one of the prompts. It was a single line: “summarise this CV”. I rewrote it together with her in about ten minutes. By the end of that session, the same model was producing screening notes that genuinely sped up her work.
That gap, between an AI tool that feels mediocre and one that feels useful, is almost always a prompt-engineering gap, not a model gap. The discipline behind it is what you and your team are paying for, whether you call it that or not.
What is prompt engineering?
Prompt engineering is the practice of writing structured instructions to an AI model so it produces consistent, useful output. In plain terms, it is the difference between typing “summarise this” and typing “you are a recruitment screener for a UK fintech firm; summarise this CV in three sentences, flagging FCA-relevant qualifications and noting any stated compensation expectations”. Same model. Same kind of work. The structure of the instruction is what makes the output usable.
Two pieces are worth naming. The system prompt is the standing instruction set that defines the AI’s role, tone, rules, and constraints. The user prompt is what someone types in the moment. Enterprise AI tools commonly let you configure the system prompt once and have users only see the user prompt, which is where the operational discipline lives.
The shift from 2024 to 2026 is that prompt engineering has stopped being a specialist job and become an everyday operational skill. The “prompt engineer” job title that briefly attracted six-figure salaries has largely dissolved. The practice has not.
Why it matters for your business
The first reason it matters is impact. Industry benchmarks suggest that better prompting closes the bulk of the gap between “AI is unimpressive” and “AI is useful” for roughly 70 to 85% of standard business tasks. That is the largest, cheapest improvement available to a service-led SME using off-the-shelf AI. It costs nothing beyond a few hours of practice per team member, scales across every tool you use, and compounds over months.
The second reason is cost. A poorly written prompt makes the model do extra work, generate extra tokens, and return responses you have to rework. At scale, that adds up. A well-written prompt produces shorter, more usable output on the first attempt, and an AI tool’s running cost falls accordingly. Token-based pricing means inefficient prompts are an invisible tax on your AI budget that better prompting removes.
The third reason is governance. Your system prompts are intellectual property. They encode your brand voice, your compliance rules, and your decision logic for any task the AI tool handles. The questions an SME owner should be asking are: who owns these prompts, how are they versioned, what stops them quietly drifting over six months, and what happens to them when the team member who wrote them leaves. It is common for SMEs to have not thought about this yet.
Where you will meet it
You will meet prompt engineering language in vendor pitches as a reassurance phrase. “No training required, just plain-English instructions” is the structural opener, designed to address the SME owner’s worry that AI tools require a data science team. The claim is broadly true and slightly misleading: plain-English instructions work, but writing good ones is a learnable skill, not an effortless one.
You will meet it in implementation partner offerings, where prompt engineering is sometimes packaged as a separate paid service. There is a real version of this and an oversold version. The real version is when a partner sets up your system prompts for a high-volume workflow, defining the rules, the format, the escalation paths, and a versioning approach. The oversold version is being charged a thousand pounds for what your senior associate could write in an afternoon with a guide.
You will meet it inside your own team in the form of prompt drift. A useful prompt gets copied around, edited slightly each time, ends up living in three different forms in three different documents, and quietly stops doing the job. Operational teams in 2026 commonly either have a prompt library nobody reads or no library at all. The fix is governance, not more clever prompting.
When to ask about it, when to ignore it
Ask about it when you are deploying a tool that will handle the same kind of task hundreds or thousands of times: customer service responses, document summarisation, ticket routing, screening, scheduling. In those cases, the system prompt is doing real work, and getting it right multiplies across every interaction. The questions are who owns it, where it lives, how it is updated, and how you will know if it stops working.
Ignore the offer of paid prompt engineering services for a small, low-volume workflow that one team member runs occasionally. That person will learn faster by drafting and revising than by hiring a consultant. A short internal guide and a few hours of structured practice closes the gap.
There is also a ceiling worth flagging. When a vendor tells you “we’ll fine-tune the model to make this consistent”, ask first whether better prompting would do the job. Often it will. Fine-tuning is the right answer when prompting genuinely cannot deliver, when output format must be locked, when domain knowledge is genuinely missing, when volume is high enough that per-request cost dominates, but not when the prompt is simply underwritten.
Related concepts
System prompt is the standing instruction that defines an AI’s role and rules. User prompt is what someone types each time. Most production systems combine the two, with the system prompt invisible to end users.
Few-shot prompting is the technique of including two or three worked examples in your prompt so the model can pattern-match its output to your standard. Often the cheapest way to lift consistency without any technical change.
Chain-of-thought prompting tells the model to work through a problem step by step rather than jumping to an answer. Useful for analysis tasks; expensive in tokens; sometimes worth it.
Retrieval-augmented generation, RAG, is the next step up when better prompting cannot give the model the context it needs because the relevant information lives in your documents.
Fine-tuning is the step beyond that, when you have enough volume and clean data that retraining the model itself starts to pay back.
Prompt versioning, prompt governance, prompt libraries, all of these are catching up with the operational reality that prompts are now business logic. By 2027 they will be a normal part of how teams manage AI tools. SMEs that have any of this in place in 2026 are already ahead of the field.
The honest read on prompt engineering in 2026 is that it is an everyday discipline that earns its keep across every AI tool a team uses. A vendor selling you a workshop is selling you a head start, not a secret.
