AI in construction bidding: where to retain control, and when to relinquish it

Over half of UK businesses in information and communication report using AI. In construction, the figure is 13%.

That gap is from the ONS, and the numbers are surely conservative (Artificial intelligence in UK businesses: 2023 to 2026). What it tells you is that construction is struggling to embrace the possibilities inherent in the technology. A traditionally archaic sector is presently locked into the mindset that AI is a sophisticated search engine, rather than something that can scan masses of information and return a concise examination of the contract risks across thousands of pages — along with the places where the documents clearly contradict themselves.

And the part that nobody selling you an LLM skin on a .ai domain will tell you: the tools are the easy part. The skill that matters is knowing which work to hand over and which to keep. That judgement is entirely human, it cannot be bought as a subscription, and it is the whole of what follows.

What that looks like on a live bid

Two decisions from recent work, in opposite directions.

On a recent bid, the tender pack included a substantial set of Employer's Requirements documents - nine individual documents and more pages than I could honestly bill a client for reading. Much as I would love to read and understand every one of them, the activity would have meant an unjustifiable amount of time at the client's expense and a total cost that would likely never be signed off. I took a local copy and had Claude summarise them into a concise and useful output, which produced a slew of clarification points. Those went in to the client and drove the mid-bid questions.

The same bid needed a large set of team CVs, with specific information requested for each person. That stayed with me. An LLM can tell you what needs to be included and give you a format for gathering it. What it does not understand is why a project manager will struggle to complete the task if it is laid out directly, exactly as the tender documents demand. Building a template that is clear and genuinely completable by the person filling it in is a slightly administrative but vital piece of work, and it is retained by the trusted bid advisor every time.

Carefully targeted effort is the principle. Everything below is in service of it.

A curated tech stack - with no obvious or easy answer

I use Claude and Gemini, for different tasks, plus Granola, Lovable, Wispr Flow and Apollo on a daily basis. All either entirely or partly powered by AI. I worked out which tools and platforms I needed first, and then tested extensively to understand which were best for my specific way of working.

For meeting notes - in person and online, platform-agnostic - I found Otter a little fiddly, but loved the simplicity and fidelity of Granola. That is not any form of recommendation. It is just what is true for me. To make tech work for you, there is a very human piece of work to do at the front end which will define what you use into the future.

As part of this testing you will likely also realise that, in areas where you thought you needed assistance, you actually prefer to handle the task yourself. One of the least imaginative suggestions on using automation is to have an LLM scan and action items in your email inbox. I hated this, as I prefer to read, understand and action things as I see fit — but of course that will not be true for everybody. I only learned by doing.

It is worth saying also, because I tried to avoid it for a while myself, that you need to accept the need for investment in a certain number of subscriptions. That makes the assessment period all the more important. There are usually free versions of most things, but in avoiding paying any money you will end up making your working life harder. Work out a budget and pay for your chosen platforms. Thank me later.

Learn prompt engineering

It is a hideously overblown term for 'providing full and clear instructions', but it is an absolute necessity. The difference between an open-ended question and one that is structured, clear and establishes the correct guardrails within which the LLM must operate is absolutely enormous.

Models produce a plausible answer whether or not they have enough to go on. So:

• Tell it to summarise a given document, and it will make assumptions based on the information at its disposal - including, where possible, the way you have behaved in the past - to return a response. Which means your summary will most likely be based on what other people like to see, rather than what you actually need.

• Instruct it to summarise the document as part of fulfilling a specific role, with the purpose, the desired ways of analysing the information, the outcomes sought and the actual output you want delivered, and you will receive a response that is useful and relevant. It will almost certainly tell you something you had not realised to be true, and will largely eliminate the possibility of 'hallucination' - the machine enthusiastically filling in the gaps, then straying into areas where it gets things wrong.

As part of the learning I invested in, I was introduced to a completely free prompt improver which does a lot of the heavy lifting for you. If you take one thing away from this, make it this extension (ivee Prompt Improver).

Understand where to retain control, and when to relinquish it

The temptation, because we are human and therefore inherently prone to being deeply lazy, is to devolve responsibility for as much of our daily work to AI as possible. It is an understandable temptation. If you can make your life easier by using various tools, why wouldn't you?

The answer is that there are things AI cannot replicate in how humans perceive behaviour and adjust an approach to get the right result. The CV template above is one of them. The Employer's Requirements summary was not, and the difference between the two is not about the technology at all - it is about who the output is for, and what happens to it next.

Anything that will be read by a person you need to persuade, or completed by a person you need cooperation from, keeps a human in it. Anything that is you processing information in order to form a view can usually be handed over, provided you check what comes back.

What does all this mean, in practice?

We are now living in an era where we evolve with technology and adapt to its use, or begin to fade into the background. An unpleasant truth in many ways, but one that stands.

In the past, you would receive a response to a question after the PM had spent ages rummaging about in the shared folder for something similar, likely changing a couple of the project names near the start, then sending it across with the advice that you should "feel free to change it if you need to" - in other words, telling you to rewrite it at the point you realised it had been copied from elsewhere.

Now you will more than likely receive a poorly instructed AI response: the product of Copilot having bounced around within the company SharePoint, returning something that creates more problems than it solves, from a person who does not understand the detail well enough to have any idea whether the response is brilliant or completely useless.

---

We work across public sector frameworks including GCA, Pagabo, SCF, Fusion21 and NHS SBS - advance intelligence, framework strategy, answer planning and full bid management. See how we work on framework procurements.