AI Brand Strategy: Positioning Workshops and Messaging for Agencies

An agency can use AI in brand strategy to transcribe and synthesise stakeholder interviews, draft three to five positioning options from that evidence, and build a message house from the option the client picks. People still decide which evidence counts, which option to bet on and which claims the client can prove. Present options with their trade-offs, never a single answer.

Vendor tools tend to build a strategy from one person's answers, usually the owner's: one, for instance, talks the owner through a framework in a voice conversation and turns it into a slide deck. An agency's strategy rests on what many people say, which is why the interviews matter more than the drafting. The project below follows a home-care agency through twelve interviews, a half-day workshop and a client presentation. Every number in it is illustrative.

Follow me on Instagram@sagnikteaches

Before interview one: consent, recording and where transcripts live

Interviews work because people say things they would not put in a survey, so the first job is protecting them. Recording rules differ from place to place, and a client's own contract may add confidentiality terms of its own. The safe routine is the same everywhere. Tell every interviewee in advance that the call will be recorded and transcribed by an AI tool, say who will see the transcript, and ask for an explicit yes at the start of the recording.

Connect on LinkedInSagnik Bhattacharya

Then decide where the transcripts will live. Use a business plan, which for ChatGPT Business, Claude Team and Gemini in Workspace does not train on your content by default, or a personal account with the training setting switched off. Both ChatGPT Business and Claude Team need at least two seats, so a one-person consultancy may be on a personal plan. Strip names, client names and anything identifying before a transcript goes anywhere near an assistant. Keep the key that links "Interviewee 7" to a real person in a separate file that no assistant ever sees.

Subscribe on YouTube@sagnikteaches

The home-care agency in this example has about 60 carers and 220 clients, and its owners want to be known for something sharper than "quality care". Its sector adds a second reason for caution: interviewees will mention vulnerable people, so the studio agrees with the owners that no client names or health details will be recorded, and the interview guide tells interviewers to steer away from them.

Choosing twelve voices and writing the guide

Twelve interviews across four groups is a workable shape: three from leadership, three from carers and coordinators, four from clients' families and referral partners, and two from people who left or never signed up. The last group is the one clients forget, and it is often where the sharpest positioning evidence hides. Allow 40 minutes each, plus a day of scheduling, which for twelve people is about 8 hours of interviewing and 4 of chasing diaries.

An assistant is useful for drafting the guide from the project brief, and then it needs editing by someone who has done interviews. Five core questions go to everyone so that answers can be compared. Role-specific questions sit on top.

CORE QUESTIONS (everyone gets these)
1. Describe the agency to a friend in two sentences.
2. Who do we do our best work for, and who should we politely turn away?
3. Think of the last time someone chose us, or left us. What decided it?
4. Which other providers do you compare us with, and what do they do better?
5. If we disappeared next month, what would you miss, and what would you not?

EXTRA FOR REFERRAL PARTNERS
6. What makes you comfortable handing a family over to an agency?
7. What have you seen an agency promise that it could not deliver?

Question 5 is the one that tends to produce positioning, because answers to "what would you miss" are the benefit in the customer's own words. Question 7 is the one that protects the client: it tells you which claims the market has stopped believing. If you were tempted to skip interviews and ask an assistant to invent personas instead, are AI-generated customer personas accurate? explains how to test what a persona is really made of.

Transcribing, cleaning and anonymising 12 interviews

Transcription is the cheapest win in the whole project. Use the transcript feature in your meeting platform if it has one, or a dedicated note-taker. Otter's Pro plan, for instance, lists $8.33 per user per month billed annually or $16.99 month to month, with 1,200 minutes a month and a 90-minute cap per meeting (as of November 2026). Twelve interviews of 40 minutes use 480 minutes, so one month of Pro covers the whole project. Note the free plan's limits before you rely on it: a 30-minute cap per meeting would cut every interview off before the best answers arrive.

A raw transcript is not ready for analysis. Spend about two hours across all twelve on cleaning: fix speaker labels, correct the proper nouns the transcriber mangled (for care, that means medication names, local service names and the agency's own name), and replace names with codes. Ask the assistant to flag anything that identifies a person or a client so you can remove it by hand. Do not accept its redaction as final, because it will miss a nickname or a street name and you are the one who answers for it.

Synthesis that traces every theme back to a transcript

Load the cleaned transcripts into a source-grounded tool or a project. Gemini Notebook, the product formerly called NotebookLM, answers only from the sources in a notebook and points to the passage behind each answer; on a free account a notebook holds up to 50 sources, so twelve transcripts and five competitor PDFs fit comfortably. A ChatGPT or Claude project also works, but ChatGPT limits the number of files in a project (five on the free plan and 25 on Plus at the time of writing), so merge transcripts or check your plan's limit first. Whichever you use, the rule is the same: no theme without transcript IDs and a verbatim quote.

Using only the 12 interview transcripts (I1 to I12), identify the 8 most common themes about why clients choose, stay with or leave the agency.
For each theme give: a one-line label; the interview IDs that mention it; one verbatim quote with its ID; and any interviewee who disagrees, with their quote.
Then list: (a) the 20 most distinctive phrases interviewees use for the agency, with IDs; (b) anything important that NOBODY mentioned.
Do not paraphrase quotes. Do not count a theme unless the interviewee raised it unprompted or in answer to question 5.

The output is a theme table, which the studio then checks. Here is the top of it from the care agency project, with sample quotes invented for this illustration.

Three checks stand between this table and the client. First, find every quotation in its transcript: assistants sometimes attach a perfect quote to the wrong interview, and on this job one striking line about "never a stranger at the door" turned out to come from a different interviewee than the one cited. Second, recount the top five themes by hand from the ID lists, because the assistant's "9 of 12" can include an interview that only touched the theme in passing. Third, read at least four whole transcripts yourself: the most senior voice, the most critical, one front-line carer and one lapsed client. A summary shows you what was common; only reading shows you what was said with feeling.

The last line of the prompt, asking what nobody mentioned, earns its keep. Here the answer was that no interviewee brought up the agency's inspection rating or its training, which both owners had assumed was their main selling point. That is a finding for the client, and an assistant is good at spotting a gap like that once it is asked directly. The wider method for turning notes into a deck is covered in interview notes to a findings deck.

Three positioning options, scored by the assistant and then by you

With the evidence grid and a competitor claims table in the project (the table from a brand audit works well), ask the assistant for options. Give it a template so the options are comparable, and give it the evidence so it cannot fall back on stock phrases.

Using the evidence grid and the competitor claims table only, draft 4 positioning options for the agency.
Use this template for each: For [target client or referrer] who [need], [agency] is the [category] that [main benefit], unlike [alternative], because [reason to believe].
For each option also give: the interviews that support it (IDs); how many of the 5 competitors make a similar claim; what the agency would have to do for the claim to be true; and what the agency gives up by choosing it.
Avoid these words: quality, compassionate, dedicated, tailored, trusted, excellence.

That template is the familiar positioning-statement format; what makes this version useful is the extra fields. "What would have to be true" forces the deliverability question into the open, and "what it gives up" makes each option a choice. The assistant returned four options, and the studio dropped one that was just a blend of the others. Three went into scoring.

• Option 1, same faces every visit. Each client has a named team of up to three carers. Nine of twelve interviews support it, and none of the five competitors promise it.
• Option 2, someone always answers. Reachable day and night. Seven interviews support it, and two competitors already claim round-the-clock cover.
• Option 3, built for hospital discharge. Specialist transition care for referral partners. Five interviews support it, all from referrers, and one competitor makes a similar claim.

The scoring grid uses five criteria with weights, because evidence and distinctiveness matter more to a positioning bet than emotional pull. The assistant produces first-pass scores from the evidence. You then adjust them, and write down why.

The human adjustment is where the strategy happens. In the review call, the agency's operations manager explained that a named-team promise needs rota software the agency has not bought. The strategist cut Option 1's deliverability from 3 to 2, which moved its total from 49 to 47. It still came first, but the recommendation now carried a condition: do not advertise the promise until the rota can keep it. An assistant working from interview transcripts could not have known that, and a client that advertises a promise it cannot keep pays for it in complaints and lost referrals. Write down every adjustment and the reason, because that record is what you will defend in the workshop.

The message house, filled in for a care agency

A message house turns the chosen positioning into something staff, designers and copywriters can all use. The roof is the one sentence the whole organisation should be able to say. Three pillars hold it up, each with proof points beneath. The foundation holds the voice rules and the list of words never to use. Audience variants sit alongside, because a family, a referrer and a job candidate need the same promise in different words.

The proof-status column is the part most message houses lack, and it is the part that protects the client. Each proof point is marked as existing, needing measurement, or not yet true, and nothing in the last two groups goes into public copy until it moves to the first. The assistant drafts the house from the chosen option, the evidence grid and the voice rules, then the studio checks it line by line.

Build a message house for Option 1 using the evidence grid and voice rules only.
Give: a roof sentence of 10 words or fewer; three pillars of one sentence each; two proof points per pillar, each tagged EXISTS, NEEDS MEASURING or NOT YET TRUE based only on what the interviews and the client told us; a foundation with voice rules and a banned-word list; one headline each for families, referral partners and job candidates.
Do not invent figures, awards, ratings or promises. If a proof point is missing, write "gap" instead.

Expect three kinds of fault in the first draft. The assistant slips banned words back in, which is why the banned list travels with the prompt; the competitor frequency table described in building a banned-phrase list for a brand is a good source for it. It adds claims nobody supplied, such as "available 24 hours a day" or "award-winning", which is why the prompt forbids invented proof. And it writes in a smooth, generic register that none of the interviewees used. Replace those lines with phrases from the distinctive-phrases list in the synthesis step, and use the guidance on keeping brand voice consistent in AI-written content to turn the voice rules into something a model can follow.

Apply one swap test to the roof and each pillar. Replace the agency's name with a competitor's. If the sentence still works, it is not positioning, it is a category description. "Quality care you can trust" survives the swap. "The same few carers, every visit" does not, because only an agency that organises its rota that way can say it.

Running the positioning workshop with an assistant in the background

The workshop is a half day with the two owners, the operations manager and one coordinator. Its job is to choose, not to brainstorm, so the studio arrives with the three scored options and leaves with one decision and a list of conditions. The assistant works before and after the session, and quietly during it. It does not facilitate and it does not vote.

The silent vote comes before the discussion for a reason. Once the most senior person speaks, everyone else tends to drift to her view, and you lose the very disagreement you came to surface. In this workshop the owners voted differently: one for Option 1, one for Option 3. That split was the most useful moment of the day, because it exposed that one owner thought of the agency as a family brand and the other as a specialist partner to hospitals. The room spent twenty minutes on that question and left with a single answer.

The decision record is a half-page document the assistant drafts from the transcript and a person corrects: the option chosen, the votes before and after discussion, the trade-off accepted ("we will not compete on round-the-clock reach"), the conditions (rota software, answer-time measurement), and who owns each. Send it the same day. A decision that lives only in people's memories gets reopened in the next meeting.

Six points where the strategist decides

Everything above can be partly delegated. These six points cannot, and being able to say so is part of what clients are buying.

1. Which evidence carries weight. Nine of twelve interviews outweigh an owner's favourite theory, but one lapsed client's reason for leaving may matter more than three satisfied ones. The assistant counts; you weigh.
2. What silence means. If nobody mentions the inspection rating, is it unimportant or simply assumed? That judgement comes from knowing the sector and listening to tone, which a transcript flattens.
3. The trade-off. Every option gives something up. Choosing which loss the client can live with is a conversation, not a calculation.
4. Whether a claim is true. Only the client can tell you whether the rota supports a named-team promise. Nothing goes live while its proof status is "not yet true".
5. Who might be harmed. In this sector, a promise that families rely on is more than marketing. Check wording that touches safety, health or vulnerable people with the client's compliance lead.
6. How to handle uncomfortable quotes. A carer's criticism of management is valuable and risky. Decide with the client how to present it so that the person is not identifiable.

Presenting to the client: evidence, options and the trade-off

The presentation to the wider leadership team has one job: let people who were not in the workshop see why the decision is sound. Order the story so that the evidence comes first and the recommendation arrives as a conclusion, not a surprise.

1. What we heard. Five themes, each with its count out of 12 and one anonymised quote. Nothing identifiable.
2. What the market says. The competitor claims table and the white space it leaves.
3. Three options. One slide each, using the positioning template, with the trade-off in a box.
4. The scores, and the adjustments. Show the weighted grid and the one change made after the operations call, with the reason.
5. The decision. The option chosen, the votes before discussion, and what the client agreed to give up.
6. The message house. With the proof-status column visible, so the room sees which claims can go live now.
7. What must be true in 90 days. Rota software, answer-time measurement, weekend cover, each with an owner and a date.
8. What comes next. If the agency needs a new name or a new identity to carry the position, the next stages are the naming workflow with domain and trademark checks and the identity work.

Two habits keep the presentation honest. Never put an assistant's slide on screen unedited: its summaries read well and quietly lose the numbers. And keep the proof-status column in the message house slide even though clients sometimes ask to remove it. It is the slide that shows you were careful, and it is the one that a compliance lead in a care business will read first.

Hours and tool costs for twelve interviews

These planning figures are for the care agency project, and your own timesheets will replace them after a couple of jobs. The biggest saving is in transcription and synthesis, which is also where the biggest checking cost sits, so the saving is real only if you keep the verification steps.

Costed at an illustrative $110 an hour for strategist time, that is about $4,675 against $8,195, a difference of $3,520 over the project. The software bill for the month is small: Otter Pro at $16.99 on a monthly plan and an assistant plan at around $20, so under $40 in all. Interview consent, careful checking of every quote and a person making each of the six decisions are what stop that saving turning into a liability. Used that way, the assistant shortens the days spent on transcripts and tables, and leaves the days that matter, the interviews and the decision, to the strategist.

Questions agencies ask about AI in brand strategy

Can synthetic personas replace stakeholder interviews?

No. A persona written by a model reflects what the model has read, not what this client's customers and staff think. Use AI to synthesise interviews you have actually held, and use generated personas, at most, as a prompt for writing better interview questions. Positioning built on invented people is positioning nobody can defend in the room.

How many interviews are enough?

Ten to fifteen across four groups (leadership, delivery staff, customers or referrers, and lapsed or lost customers) is a workable range for a small or mid-sized client. Stop when new interviews only repeat existing themes. Below eight, one forceful voice can dominate the synthesis, which is exactly what interviews exist to prevent.

Is it acceptable to put interview transcripts into an AI tool?

Only if the people interviewed were told and agreed, the client's contract allows it, and you use a business plan or a personal account with training switched off. Anonymise names and sensitive details first. If you cannot say where a transcript is stored and who can read it, do not upload it.

Should the client see the assistant's positioning options as written?

No. Rewrite them in your own voice, cut the ones you would not stand behind, and add the trade-off of each. The client is paying for your judgement about which option is worth betting on. Assistant drafts are raw material, and they tend to sound alike across projects.

Further reads

• AI for Branding Agencies: Where It Fits in a Brand Project — See where strategy sits between research, naming and identity work.
• Best AI for Brand Identity: Logos, Palettes and What Still Needs a Designer — Carry the chosen positioning into logos, palettes and guidelines.
• Best AI Tools for Product Marketing at a Small Company: Launch Kit — Turn a positioning choice into launch assets for a product.
• AI Creative Brief Generator for Agencies: From Client Call to Brief — Convert the workshop decision into a brief the team can use.
• Best AI for Consulting Presentations: Gamma, Copilot, ChatGPT — Compare tools for building the client presentation.
• Best AI for Consulting Research: Market Scans With Source Checks — Check market claims before they reach a strategy slide.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Otter's pricing page, a strategy-tool vendor page, Google's Gemini Notebook help pages, OpenAI's ChatGPT Projects help article, the ChatGPT Business and Claude pricing pages and Google Workspace's AI privacy page, all read in November 2026.