AI for agency new business is using models to compress the slow, repetitive work around a pitch — researching the prospect, drafting the response, tailoring credentials, assembling the proposal — while the relationship and the creative idea stay with people. Agencies that use it well win more time for the parts of a pitch that actually decide it; agencies that use it badly produce more generic pitches faster and lose.
New business is where an agency's growth story is won or lost, and it is also one of the most repetitive parts of running an agency. That combination makes it a good candidate for AI, provided you are honest about which half of the work is actually repetitive.
What new business actually costs
The cost of new business is not the meeting; it is everything around it. Reading the prospect, understanding their category, tailoring the case studies, writing the response document, building the deck, chasing the follow-up. Most of that work is structured and repeatable, and a large share of it is done under time pressure by people whose real value is judgement and relationships.
That is the seam AI fits into. It removes the assembly that stops senior people from thinking, so the pitch gets more of their best hours. It does not write a better idea, and any agency that outsources the idea to a model will be found out the moment the client asks a real question.
It is also worth remembering that a pitch is partly an argument for why the work should not simply go in-house. That calculation is old — I wrote about it when search marketing was moving in-house — and AI has not changed its shape: the agency has to show a standard and a memory the client cannot build alone.
Where AI helps across the pitch
Four stages suit models, and each has a clear human counterpart that must not be given up. The table below is the split I would draw.
| Stage | AI's job | The human's job |
|---|---|---|
| Prospect research | Summarise public material, map the category | Decide what matters to them |
| First draft | Assemble structure and boilerplate | Write the argument and the idea |
| Credentials | Tailor existing case studies | Choose which proof is relevant |
| Proposal | Format, check, version | Own the numbers and the promise |
The pattern is that AI handles assembly and the human handles selection and commitment. When the split blurs, you get pitches that are polished, complete, and indistinguishable from the last ten the client saw.
Research without fabrication
The first place AI earns its keep is research. A model can summarise a prospect's public material, map their category, and surface the questions a competitor might ask, in an hour rather than a day. That is genuinely useful, and it is also where the danger lives, because models will invent facts about a prospect with total confidence.
The rule is simple: use AI to find and organise, and verify anything that goes into a document a client will read. No invented metrics, no assumed relationships, no "the prospect said" that you cannot source. A pitch built on a fabricated detail is a relationship-ending risk, and the effort saved is not worth it. Keep a source for every claim that matters.
Draft, but never send the draft
Writing the response is where AI is most tempting and most dangerous. A model can produce a competent first draft that covers every section, and if you send that draft you have sent a generic pitch. The value is in using it as scaffolding — structure, coverage, a checklist of what a response should contain — and then writing the actual argument yourself.
The practical rule is that no AI-authored sentence goes to a prospect unedited. Not because the prose is bad, but because the pitch is a promise, and a promise you did not write is one you cannot stand behind. The editing time is where the agency's point of view gets inserted, and it is the part that wins work.
Credentials without inventing
Tailoring credentials is a task AI does well when the raw material is real and a liability when it is not. Feeding a model your actual case studies and asking it to select and sequence the relevant ones against a prospect's brief is legitimate and fast. Asking it to "write a case study" is how agencies end up with proof they cannot defend.
Keep a structured library of real case studies — the problem, the work, the result, the evidence — and let the model work from that. Over time this library becomes an asset in its own right, and it is the same memory layer that underpins the rest of an AI-native agency. The richer and more truthful it is, the more useful every future pitch becomes.
The generic-pitch trap
The greatest risk is not that AI makes a bad pitch; it is that it makes a competent, generic one. A model trained on everyone's pitches produces the median pitch, and the median pitch does not win. If your new-business output starts to feel smooth and interchangeable, the problem is not the model; it is that you let it write the parts that should have been yours.
The counter is deliberate: use AI for everything except the idea and the point of view, and make sure a senior person owns those. The test is whether a prospect could tell your pitch apart from a competitor's if the logos were removed. If not, the AI saved you time and cost you the work.
A new-business workflow
Build the workflow in four steps. Collect the prospect material and have the model summarise it. Draft the response structure from that summary and a template. Select the best real case studies from your library and tailor them. Then have a human write the argument, check every claim, and own the numbers. Automate the first three steps and treat the fourth as the work.
Over time, instrument the workflow the way you would any other: how long a response takes, how much senior time it consumes, and what the win rate looks like. That is the honest way to know whether AI is helping new business or just making the paperwork faster. The agency-model guide covers the wider changes AI brings to how an agency runs.
The follow-up and the pipeline
The pitch is not the only place AI helps; the follow-up is where most agencies lose momentum. A model can draft the follow-up sequence, summarise the meeting, and keep the pipeline notes current, so the relationship stays warm without someone manually chasing every thread. None of it changes who owns the relationship, and none of it should automate the personal note.
The habit that matters is turning each pitch into a learning. Win or lose, capture why, add it to your case-study and positioning library, and let it inform the next pitch. That library is the new-business equivalent of the memory layer, and it is the reason a mature agency pitches faster and better than a young one. Feed the model from it, and let it compound with every conversation. A short written note after each pitch — what was asked, what landed, what you would do differently — takes minutes and is worth more than any tool, because it is the raw material the next pitch is built from. That note is also what makes the model useful: without it, the AI is drafting from nothing.
FAQ
Can AI write our RFPs?
It can draft the structure and the boilerplate, and it should not write the argument or the idea. Use it to make sure you cover every section and to save assembly time. The words that carry the pitch, and the commitments inside them, stay with a person who can defend them.
Is it ethical to use AI in a pitch?
Yes, when you use it for research, drafting, and assembly — the same categories of work account teams have always delegated. It becomes a problem when you present fabricated proof or pass off a generic response as bespoke thinking. The rule is the same one you would apply to any junior support work: it must be true and it must be yours.
Won't every agency use this and cancel the advantage?
They will use it, which is why the advantage moves to the parts models cannot do: the idea, the relationship, and the real case studies. When everyone has fast assembly, faster assembly stops being a differentiator. The agencies that win will be the ones whose thinking and proof are genuinely better.
What is the biggest mistake?
Sending the first draft. It is the fastest way to lose a pitch, because the prospect can feel the absence of a point of view. Use the draft as scaffolding, then write the pitch yourself, even under time pressure. That editing step is the job.
The short version
AI compresses agency new-business research, drafting, credential tailoring, and proposal assembly so senior people spend their time on the idea and the relationship. Use it to find and organise, never to fabricate, and keep a human on every sentence a prospect will read. The agency-model guide covers how the model changes, and the for agencies track works new business into the agency operating model.
If you run new business, I would like to hear where you draw the line between what AI drafts and what you write yourself.
Cheers, Chandler