AI changes the preparation. The room still belongs to a person.

AI-assisted preparation means using tools such as large language models or research assistants to pull together background on a company, its market and its people before you sit down to write anything. Done well, it turns a task that used to eat an afternoon into one that takes twenty minutes, and it gives you room to spend the time you saved on the parts that actually need judgement: what this buyer cares about, and why.

What to automate

The compiling work is where AI earns its keep. Recent financial results, leadership changes, competitor moves, prior suppliers, published strategy documents and industry commentary can all be gathered and summarised faster by a research tool than by a person working through search results one tab at a time. This is background reading, so the risk of getting it wrong is low and the time saved is real.

What to check before you use it

• Whether the facts are current. A model trained or indexed some months ago can state a company's old leadership team, old financials or a closed office as though they still hold.
• Whether the facts are real. Language models produce plausible-sounding detail that did not happen, sometimes called hallucination. A confident sentence about a client's award, contract or partnership is worth checking against a primary source before it goes in a slide.
• Whether the source was personal data. Scraping or summarising named individuals' LinkedIn activity, job history or personal posts for a pitch touches UK GDPR, and business use of that data should be something you could explain to the person concerned if they asked.

A single wrong fact about the buyer's own business does more damage than an hour of missing research would have. Buyers notice when you've got their numbers wrong far more readily than they notice when you've done extra homework, so the check is not optional.

Where the edge is real

The advantage is not that you used AI. It's that the time it freed up let you go further into the specific situation: the stakeholder who blocked the last vendor, the internal metric this team is actually measured on, the question nobody on the buying side has been asked yet. That depth is what a buyer credits to you, because they never see the tool.

Where it turns into a liability

Preparation that reads as generic, or that repeats information the buyer already knows back to them with new phrasing, signals that the research was done to a template. The same applies if a stakeholder can tell their profile was compiled by a machine, most often because the detail is accurate but the framing is flat. AI can find the facts. Deciding which three of them matter to this buyer, in this room, is still the work only a person can do, and that judgement is described on the about page as the standard this site holds its own research to.

A proposal written largely by an AI model reads differently from one written by someone who sat in the room, and buyers who see enough of them learn to notice the difference within a paragraph or two. The tell is not that the writing is bad. It is often smoother than a rushed first draft would be. What gives it away is the absence of anything the model could not have known: no reference to the objection raised on the call, no number pulled from the buyer's own numbers, no phrasing that only makes sense if you were there.

What the tells actually look like

The specific signatures repeat often enough to be recognisable. None of them proves a proposal was AI-written, but together they are a strong signal to an experienced reader.

• Generic framing that could sit in front of any client in the sector, with no detail specific to this one
• Uniform paragraph rhythm, where every section runs to a similar length and structure regardless of how much the point actually needs
• Hedge language such as "can help drive" or "is designed to support", which commits to nothing and can be checked against nothing
• Inflated adjectives like seamless, robust or transformative doing the work that evidence should be doing
• No trace of the discovery conversation, so the document reads as though it was written before the call.

Why it costs more than it saves

A pitch document is not the decision. The buyer decides on trust, status and story, and the proposal supplies the reasons they use to justify that decision to themselves and to whoever else signs off. A document that reads as generated tells the buyer something uncomfortable: that the effort they can see did not go into understanding their problem. That reading does not need to be fair to do damage, because it arrives before the buyer has read past page two.

The cost lands hardest at the moment it matters least conveniently, which is when the buyer is comparing two proposals that are otherwise close. A generic document gives them nothing to defend it with in the room where the real decision gets made, and a champion who wanted to back you now has less to work with.

What to do instead

AI tools are genuinely useful for structuring a draft, checking a proposal against a brief, or turning scattered notes into a first pass. The failure mode is publishing that first pass unchanged. The fix is not to avoid the tools; it is to make sure the finished document carries something the model could not have supplied: a number from the buyer's own data, a phrase they used on the call, a judgement about their specific situation that only follows from having listened to it.

An AI note-taker joins a call, listens to everyone in the room, and turns their speech into a transcript and a summary of action points. The tool itself is not the issue. The issue is that the moment it starts listening, you are processing personal data about everyone on the call, whether they are your side or the buyer's, and UK GDPR applies to that processing from the first second of the recording.

What UK GDPR actually requires

Recording someone's voice and converting it to text is personal data processing under UK GDPR, and processing needs a lawful basis. Consent is the obvious one, but it has to be informed: the people on the call need to know, before the recording starts, which tool is running, what it does with the audio, and who will see the transcript afterwards. The ICO's guidance on recording and monitoring treats a quiet assumption that everyone is fine with it as insufficient. Silence is not consent.

PECR (the Privacy and Electronic Communications Regulations) governs electronic marketing communications and cookies, and it is not the operative law for recording a meeting you are already in. Do not let a passing reference to it stand in for the UK GDPR analysis above: the transparency and lawful-basis obligations sit with UK GDPR, and treating PECR as the relevant rule here would be a distraction from the one that actually applies.

Where this goes wrong in practice

The common failure is not a legal one. It is a trust one. A transcript captures everything said, including the aside your colleague made about the buyer's procurement process, or the moment you floated a number you had not cleared with your own partner. Once a transcript exists, it is discoverable and forwardable in a way a memory never was, and a buyer who later sees it circulated without their knowledge will remember that more clearly than anything in the pitch itself.

There is also a negotiation cost specific to your side. A transcript is a record of your walk-away position (your BATNA, or best alternative to a negotiated agreement) if you say it out loud, and of every number you tested before landing on the one you meant to open with. Recording tools do not distinguish between the rehearsal and the real offer.

The practical rules

• Announce the tool by name before it starts: say what it is and what it will produce.
• Get an acknowledgement from everyone on the call.
• Pause or stop it for informal exchanges you would not want written down verbatim.
• Check the tool's data retention and model-training terms before the meeting.
• Store the output where your firm controls access.

None of this replaces legal advice on your specific processing arrangements, and where a deal touches regulated activity, such as advice on financial products, insurance, credit, legal or medical services, that advice remains the domain of the FCA, the SRA or the relevant regulator. The starting point for the general rules is the ICO's own guidance on recording conversations and monitoring communications.

A remote pitch used to carry an unspoken assumption: if you can see and hear the other person on the call, you know who you are dealing with. That assumption no longer holds. Voice cloning and video deepfake tools have moved from research demonstrations to tools that can convincingly impersonate a named individual on a live call, and buyers who have read about the failures are starting to build verification into how they take meetings.

The mechanism, illustrated

In 2024, an employee at the engineering firm Arup authorised a transfer of roughly HK$200 million after a video call with people who appeared to be the firm's own UK-based finance staff. Hong Kong police later confirmed the entire call, faces and voices included, had been generated. The case is worth knowing because it shows the failure point clearly: the deception happened inside a video call that looked and sounded exactly like the real thing.

What buyers are starting to check

Procurement teams and finance functions who have seen cases like this circulate are adding low-friction checks to remote pitches, particularly where the meeting will lead to a payment, a contract signature, or a change of bank details. None of these prove identity beyond doubt, but each raises the cost of faking it.

• A callback to a number already on file
• Confirmation of any changed payment details through a separate channel, such as a phone call to a known contact, before acting on them
• A request for an unscripted action on camera, such as turning to answer a question about something not on the agenda, which current deepfake tools handle poorly under real-time pressure
• Cross-checking the meeting invite against the counterparty's own corporate domain

Where verification becomes theatre

These checks work against opportunistic fraud, and treating a callback or a live-camera request as proof of identity is a mistake in itself. The more common failure is the opposite one: firms bolting on visible verification steps that reassure nobody and slow down every legitimate pitch, while doing nothing about the accounts and calendars that were never checked in the first place. The steps above reduce risk; they do not remove it, and no verification step should be presented to a buyer as a guarantee.

None of this changes who is accountable for a pitch. The person presenting still owns what is said in the room, deepfake risk or not, and the approach behind this site treats verification as one more thing a serious pitch process now has to plan for, alongside the argument itself.