How to Create a Sales Battlecard Using AI That Survives a Live Call

Why AI Battlecards Fail Mid-Call

AI battlecards collapse under pressure for one structural reason: a single prompt produces a document that looks finished without any of its claims being tested. The model writes just as fluently about a feature it half-remembers as about one it can cite, and on a page of bullets, fluency is indistinguishable from accuracy.

This is not the model being careless. It is doing what it was asked. You requested a battlecard, so it produced the shape of a battlecard, and it filled the gaps in that shape with plausible-sounding content. The mechanism behind that gap-filling is worth understanding properly — our explainer on why AI invents facts and how to catch it covers it in depth.

The second version is shorter, checkable, and says something a rep can actually use. The first one is not wrong so much as empty, and empty lines are what teach reps that the battlecard is not worth opening. That pattern is well documented outside AI too — Klue’s competitive battlecard guide makes the same point about manually built cards: the failure is almost never missing information, it is information reps cannot act on.

The Field-Ready Filter Explained

The Field-Ready Filter is a four-gate test every line on a battlecard must pass before a rep says it out loud: Sourced, Asymmetric, Speakable, Current. It is the approach we teach at PromptPeakAI for turning an AI draft into something you would put in front of a customer.

The order matters. Sourced comes first because it is the cheapest gate to fail — you either have the URL or you do not, and roughly a third of an AI draft dies right there. Running Speakable first would mean polishing the phrasing of claims you are about to delete.

Gather Intel Before You Prompt

Your card is only as good as what you put in front of the model, and the highest-value inputs are the ones the model cannot reach on its own. Public pages it can find. Your own deal history it cannot.

Pull three tiers of input before you open a chat window:

• Primary pages — the competitor’s pricing page, product pages, docs, changelog, security or trust page.
• Independent signal — published reviews and comparison threads, where you count recurring complaints rather than reading individual ones.
• Your own evidence — anonymised notes from deals you won and lost against them. If you have never structured these, our guide to win/loss analysis with AI is the place to start; it is the single input that makes your card different from a competitor’s.

That third tier is where most people get nervous, and rightly so. Deal notes contain customer names and numbers.

If your company has a written AI policy, read it before you paste a transcript. Our practical guide on whether ChatGPT is safe to use at work walks through what to check and what to anonymise.

How to Create a Sales Battlecard Using AI

Here is how to create a sales battlecard using AI without any of that fragility: four prompts, run in order, roughly 25 minutes. Each one produces an input the next one needs, which is exactly why a single mega-prompt underperforms: it collapses four different jobs into one, and the model averages across all of them.

Each prompt below follows the Role, Goal, Context, Format structure — if the shape looks deliberate, that is why. Our breakdown of the Role, Goal, Context, Format framework explains what each slot is doing.

Turn web search on before you send this. Without it the model answers from training data, which is where stale specs come from. The instruction that does the heavy lifting is the last one: no source, no line.

You are a competitive intelligence analyst preparing source material for a sales battlecard.

Using live web search, research [Competitor Name] at [competitor URL].

Return a factual brief covering:
1. How they describe their own positioning, in their words
2. Their pricing model and packaging structure
3. Their stated integrations and platform requirements
4. Any feature or limitation they document publicly

Format as a table with three columns: Claim | Source URL | Date of the page.

Rules you must follow:
- Every row needs a live URL I can open. If you cannot find a source, omit the row entirely.
- Do not infer, estimate, or fill gaps with what is typical for this category.
- If something is genuinely unclear from public sources, add a final row saying so.

This is Gate 2 as a prompt. The model’s instinct is to list everything you do that sounds good, most of which the competitor also does. You have to force the comparison.

Using the sourced brief above, compare [Competitor Name] against our offering.

Our product: [product name].
What we actually do differently: [list 3 to 5 specifics — architecture, pricing model, support model, deployment]
Notes from deals we won and lost against them: [paste anonymised notes]

Produce three lists:
1. ASYMMETRIC — differences where only one of us can make the claim
2. PARITY — things we both do, which must never appear as a differentiator
3. THEIR ADVANTAGE — where they are genuinely stronger, stated plainly

For the review-based items, only include a weakness if three or more independent reviews raise it. Ignore single complaints and one-off bug reports.

Be blunt in list 3. A card that pretends we win everywhere gets ignored.

Gate 3 is where most cards die quietly. A rep will not read a paragraph aloud, so anything longer than about two sentences never gets used. The word-count cap in this prompt is not decoration; it is the whole point. This mirrors the structure in OpenAI’s own sales prompt library, which pairs battlecard generation with a short talk-track requirement rather than a document dump.

You are a senior enterprise account executive coaching a new rep.

Take the ASYMMETRIC list above and write, for each item:
- One 15-second spoken response the rep can say on a live call
- One open question that lets the prospect discover the difference themselves

Voice rules, applied strictly:
- Write it as speech between two professionals, not as marketing copy
- Maximum 40 words per response
- Never name the competitor dismissively, and never say anything you would not say if their rep were on the line
- Banned words: industry-leading, best-in-class, robust, seamless, synergy, superior, world-class

Format as a two-column table: What the rep says | The question to ask instead.

Six minutes here saves the call in the opening scene. You are asking the model to argue against the document it just wrote, which it will do surprisingly well when you tell it whose side to take.

Switch roles. You are now [Competitor Name]'s top sales rep, and you have obtained our battlecard: [paste the draft card]

Go through it line by line and tell me:
1. Which claims are out of date, imprecise, or easy to publicly disprove
2. Which claims you would welcome, because you have a strong counter ready
3. Which single line would embarrass our rep fastest if challenged on a call
4. For each problem, what we would need to check to settle it

Do not soften this. Rank the problems by how much damage each would do in a live deal.

Turn Features Into Landmine Questions

A landmine question is a neutral discovery question that leads a prospect to notice a competitor’s weakness themselves, without you criticising anyone. It works because buyers discount what a vendor tells them and trust what they conclude on their own.

Three things make a landmine question work rather than backfire:

• It has to be a question you would ask even if the competitor were not in the deal.
• The answer has to matter to the buyer’s own project, not just to your win rate.
• You need a follow-up ready for the answer you did not expect.

Act as a B2B sales strategist.

Our advantage: [specific advantage, e.g. no local install required]
Their approach: [specific constraint, e.g. desktop agent per device]
Who I am talking to: [role, e.g. IT Director at a 400-person firm]

Write 4 discovery questions that lead this person to notice the constraint themselves.

Each question must:
- Sound curious rather than loaded, and be worth asking regardless of who else is in the deal
- Come with the likely answer, and my one-sentence follow-up
- Come with a second follow-up for when the answer contradicts my assumption

Never mention [Competitor Name] in any question.

If you want to build this into your standard call prep rather than treating it as battlecard-only, our guide on preparing discovery call questions using AI covers the wider question set by deal stage.

Defend Price Without Discounting

“They quoted us 30% less” is the objection that generates the most unnecessary discounts, usually because the rep has no prepared answer and reaches for the only lever they control. A three-move structure — acknowledge, reframe, redirect — gives them something better to reach for.

Notice what the second version does not do: it does not dispute the number, and it does not claim they are secretly expensive. It moves the comparison to ground where you can win it, and it does that in under 40 spoken words.

You are a negotiation coach writing lines a rep will say out loud.

The objection: "[Competitor Name] quoted us [X]% less for the same thing."

What genuinely differs in the total cost: [list — implementation fees, support tiers, add-on modules, seat vs workspace pricing, migration effort]

Write a 3-step response using acknowledge, reframe, redirect:
1. Acknowledge the number honestly, without conceding the comparison
2. Reframe to total first-year cost, naming one specific line item
3. Redirect with a question that moves us to a joint cost exercise

Constraints: under 60 words total, plain spoken English, no claim we cannot support with a document. Add one sentence on what to do if they push back again.

Where AI Falls Short on Competitive Intel

Being straight about this is what makes the rest of the card trustworthy. There are things the model cannot do here, and pretending otherwise is how teams end up with a confident, wrong document.

This is also the honest answer to the “should we just buy a competitive intelligence platform” question. The prompt workflow is not a cheap substitute for one; it is a different tool with a different failure mode.

The verification pass is the part people skip, so here is the decision it comes down to. Every claim ends up in one of three places, and “probably true” is not one of them.

Budget five minutes for this on every card. Our walkthrough on fact-checking AI research before you put your name on it covers the source-tracing method in more detail, and it applies here almost unchanged.

Fixing a Battlecard That Isn’t Working

Most bad cards fail in one of five recognisable ways, and each one traces back to a missing instruction rather than a bad model. Find the symptom, add the constraint, re-run that one prompt.

One more pattern worth naming: the card that is technically accurate and still useless because it lists parity features as differentiators. That is Gate 2 failing silently, and the only cure is running prompt 2 again with your parity list written out explicitly.

Keeping the Card Current

A battlecard has a shelf life measured in weeks, not quarters. Competitors change pricing pages without announcements, and the card does not know. So set a cadence and make refreshing it cheap enough that you actually do it.

Monthly works for an active rival, quarterly for a rare one. The refresh is not a rebuild — you are looking for what moved.

Using live web search, check [competitor URL], their pricing page, and their public changelog or release notes for the last 60 days.

Compare what you find against our current card: [paste the card]

Report only differences, as a bulleted list:
1. New features or capabilities launched
2. Pricing, packaging, or tier changes
3. Positioning or messaging shifts on their own pages
4. Which exact lines of our card are now wrong or need re-wording

For each item, include the source URL and the date. If nothing changed in a category, say "no change" rather than filling space.

If you are doing this across several competitors, the one-card-at-a-time approach stops scaling around the third rival. Our guide to building an ongoing competitor monitoring system covers the shift from periodic refreshes to a standing process.

Start with your most-mentioned competitor and one card. Run the four prompts, apply the four gates, and hand it to a rep with an honest label on which lines are verified and which are customer signal. The reason to learn how to create a sales battlecard using AI this way is simple — the card that survives its first real call is the only one worth expanding.