The AI sales stack every revenue team needs (it's smaller than you think)

Here is the AI sales stack I actually run as a quota-carrying enterprise seller in APAC: two agents I built myself — a deal-intelligence agent and an annual-report research assistant — plus two commercial tools: Gong for calls and LinkedIn Sales Navigator for outreach. That's it. Four things, and the two that create the most edge cost almost nothing.

I'm not a vendor and nobody pays me to recommend tools. I close seven-figure deals — including a US$1.2M, 18-month contract with a large bank in APAC — and I build production sales agents on frontier model APIs. This is the stack from inside the deals, not from a pricing page.

The biggest lie in AI sales content: "AI is doing the selling"

Before the stack, the thesis — because what you believe about AI determines what you buy.

The entire AI-SDR and "autonomous pipeline" category is built on the premise that the constraint in sales is activity volume. That's false for almost everyone selling anything complex. Enterprise deals don't die from insufficient emails. They die from bad qualification, no champion, no access to the economic buyer, no quantified pain. AI that 10x's your outbound volume just helps you fail faster — and burns the TAM with noise. Buyers' inboxes in 2026 are the proof: "hyper-personalized at scale" has become a synonym for spam with a first-name token.

The second-order lie is adoption theater. Salesforce's State of Sales 2026 reports that 87% of sales organizations now use some form of AI (Salesforce, 2026) — but the definition is doing heavy lifting, covering everything from lead scoring to drafted emails. Meanwhile, only 17% of sales and marketing professionals have received comprehensive, job-specific AI training (General Assembly, 2025). Leaders are buying tools to tell the board "we're AI-enabled" while behavior on the floor hasn't changed — six-figure tooling budgets producing no measurable change in how anyone sells.

Pre-call briefs, MEDDPICC gap analysis, objection rehearsal, regulatory trigger mapping — that's AI taking the four hours of grunt work before a meeting down to ten minutes, so the human shows up sharper. That's a defensible claim with measurable lift. "Our agent books meetings while you sleep" is not.

The stack, layer by layer

My stack has two layers, and the distinction matters more than any logo.

Layer 1 — the plumbing (company-bought): Gong records and surfaces what happened in my calls. Sales Navigator executes outreach. These are infrastructure: useful, standardized, and identical to what every competitor's rep is running. Plumbing doesn't differentiate you. It just stops you leaking.

Layer 2 — the edge (self-built): two agents built on general-purpose frontier models, fed my deal context. This is where the actual advantage lives, because no vendor can package what only you know about your deals, your market, and your qualification standards.

Agent 1 — The deal-intelligence agent

Built on a frontier model API, this is my MEDDPICC engine. It takes my raw deal notes — messy, mid-cycle, human — and does the work I used to do at 11pm: drafting pre-call briefs, inferring MEDDPICC gaps from what's in (and missing from) my notes, synthesizing discovery calls, and assembling forecast narratives. I've written about what each MEDDPICC letter looks like in a real deal — this is the agent that pressure-tests those letters on live pipeline every week.

The point isn't the technology. It's that the agent runs my qualification bar, on my deal data, in my framework. A generic tool can't do that, because the framework is the moat.

Agent 2 — The annual-report research assistant

This one I'll just show you. Built on Google Gems (the model matters far less than the workflow), it analyzes a target company's two most recent annual or integrated reports and produces a structured, page-referenced intelligence brief: business overview, strategic priorities and C-suite statements, sustainability reporting scope and assurance providers, material topics and who likely owns them internally, targets and ambition level, governance and board oversight, peer benchmarking, year-over-year changes — and then turns all of it into five to seven specific conversation starters for a first meeting, including two or three aimed at C-suite leaders based on their own recent statements.

Here is the actual prompt. Copy it:

You are an expert research assistant helping to analyze a company based on its
two most recent annual or integrated reports. Your goal is to extract structured,
referenced insights with page numbers, and provide conversation-starting
suggestions for engaging with the company on sustainability, governance, and
strategic matters. Analyze both documents and answer the following sections:

1. BUSINESS OVERVIEW — Legal form (public/private), listing status and
   exchanges; sector, business model and key revenue streams; key operating
   countries and total revenue (trends if available); employee breakdown
   (FTEs, types of roles); notable controversies, scandals or media mentions
   (if any). Include page numbers where each piece of info appears.

2. STRATEGIC FOCUS & LEADERSHIP — Summarize the company's strategic priorities
   and KPIs (financial & non-financial); extract the names and key points from
   the statements of the CEO, CFO, and CRO (if available); list key business
   risks and who is accountable for managing them. Include page numbers.

3. SUSTAINABILITY REPORTING SCOPE & QUALITY — Approx. number of
   sustainability-focused pages in each report; all sustainability-related
   reports published (e.g. TCFD report, modern slavery report, gender pay);
   whether a design/reporting partner is credited; whether sustainability
   disclosures are available separately online; what type of assurance is
   provided (limited/reasonable) and on what indicators; run OCR on the pages
   to name the financial audit provider and the sustainability assurance firm;
   note any "emphasis of matter" call-outs. Include page numbers for all evidence.

4. MATERIAL TOPICS & DISCLOSURES — List material ESG topics (IRO list, ESRS
   standards if applicable); call out entity-specific disclosures beyond
   regulatory requirements; suggest which internal teams might "own" these
   topics (e.g., HR, Risk, Ops). Include page numbers.

5. TARGETS & AMBITION — Extract stated sustainability goals, timelines and
   KPIs; assess level of ambition (e.g., Net Zero target by when, Scope 3
   included?). Include page numbers.

6. GOVERNANCE & OVERSIGHT — Identify Board-level and Executive-level oversight
   of sustainability; list internal committees or cross-functional working
   groups, with reporting lines; highlight whether executive compensation is
   linked to sustainability metrics; identify the internal controls or COSO
   frameworks used for sustainability data. Include page numbers for all.

7. PEER SET & BENCHMARKING — List which peer companies (if any) are referenced
   in the report (check the executive remuneration peer set); suggest
   comparable companies based on sector and model. Include page numbers.

8. YEAR-OVER-YEAR CHANGE — Highlight what changed from the previous year's
   report (added disclosures, changed KPIs, expanded assurance, new risks).
   Include page numbers.

9. INTRODUCTORY MEETING PREP — Based on the above, suggest 5–7 smart, specific
   conversation starters or questions for a first meeting with the company,
   plus 2–3 questions specific to C-suite leadership based on their recent
   statements. Focus on: the business value of action on sustainability,
   governance changes, ambition level, data/reporting complexity, or
   cross-functional alignment.

Present everything in a tabular format.

Read what that produces: by the time I walk into a first meeting, I know the company's strategic priorities in the CEO's own words, who owns sustainability at board level, whether executive comp is tied to ESG metrics, what changed since last year's report, and which questions will make a controller lean forward. That used to be four hours of reading. Now it's ten minutes of review — and the meeting opens at depth instead of discovery-by-interrogation.

The math: where the two hours come back

Across a live deal, the prep compression is worth roughly two hours of grunt work per deal, every deal: pre-call research, post-call synthesis, MEDDPICC gap-checking, follow-up drafting, forecast narrative. On a first meeting with a full report analysis, it's far more — the difference between four hours of reading and ten minutes of reviewing.

That time doesn't disappear. It moves: into multi-threading one more stakeholder, into pressure-testing the champion, into the in-person meeting that actually closes deals in this region. AI doesn't sell. It buys back the hours you sell with.

What AI must never touch

Three things stay human, permanently: the relationship, the room, and the judgment call. No agent attends my meetings, sends a message I haven't read, or decides whether a deal belongs in commit. AI drafts; I decide. The moment that inverts, you've automated your credibility.

And the question every enterprise AE silently has — what about confidential deal data? Salesforce's own research finds 51% of sales professionals say data-security concerns halt AI initiatives (Salesforce, 2026). My practice is two rules. First: I anonymize before I prompt. The model needs the deal structure — stage, stakeholder map, objection, paper process — not the customer's name. Placeholders cost me nothing and remove 90% of the risk. Second: the other 10% is tool choice — enterprise or API tiers with no-training-on-inputs terms only. If a tool can't show me that in writing, my pipeline data doesn't touch it.

How to start this week

Not with a procurement cycle. With one workflow:

1. Pick your ugliest prep task — pre-call research is the usual answer — and write down exactly what a great output looks like.
2. Build the prompt once, properly. Structured sections, evidence requirements (page numbers force honesty), and an output format you can scan in minutes. Steal mine above.
3. Run it on your next three meetings and keep score. Minutes saved, and one thing it surfaced that you'd have missed. If both numbers are real after three meetings, you have a stack. Then expand to discovery synthesis and qualification gap analysis.