Most go-to-market plans we review this year have one new line item: AI. A tool budget, a pilot with an agent, a note that the sales team will 'use Grok and Claude'. The line item is rarely the problem. The problem is that AI tools make an existing go-to-market mistake faster and cheaper to repeat, while the new openings go untouched.
Here are the mistakes we see most often, how these tools amplify each one, and the opportunities most teams have not touched yet.
Mistake one: automating outreach before the positioning is settled
The first thing most teams do with a capable model is generate more outbound. More sequences, more variants, more personalised first lines. If the positioning is unclear (who exactly this is for, what it replaces, why now), the result is a larger volume of unclear messages. Reply rates fall, sending domains get flagged, and the team concludes that AI outreach does not work.
The order matters. Positioning is a decision people have to make; volume is what tools are good at. Use the model to pressure-test the positioning first: give Claude your three strongest competitors' pages, your own homepage and ten real win/loss notes, and ask it to argue the case of the buyer who chose the competitor. That is a better use of an afternoon than a thousand new emails.
Mistake three: giving an agent a goal and no guardrails
Agents are the real change of the past year. Claude can now work through multi-step tasks with tools: read a CRM record, research the account, draft a tailored note, log what it did. That is a different thing from a chat window, and it is where most of the value comes from, and most of the damage.
- An agent with write access to your CRM and no review step will eventually overwrite good data with plausible guesses. Start read-only and grant write access field by field.
- An agent that sends without approval is speaking for your brand at machine speed. Keep a person approving outbound until weeks of reviewed drafts have gone out without needing an edit.
- An agent with no log is an agent nobody can correct. Every action it takes should be recorded somewhere a manager reads.
- An agent measured on activity will maximise activity. Measure it on the outcome the human team is measured on: qualified meetings, not emails sent.
An agent does not fix a broken sales process. It runs the broken process faster, with better grammar.
Mistake four: one tool for every job
Teams tend to pick a single assistant and route everything through it. Different tools are good at different parts of the work, and a go-to-market stack that ignores that pays for it in quality.
- Live signal, meaning what is being said about a category this week, suits a tool with real-time access to public conversation, which is Grok's home ground.
- Long, careful work over your own material suits Claude, which handles long documents and nuanced writing well: turning forty call transcripts into a messaging framework, answering a buyer's RFP point by point, holding one tone across a launch.
- Repetitive multi-step operations such as enrichment, routing and follow-up drafts suit an agent connected to your systems, with the guardrails above.
The point is not loyalty to a vendor in either direction. It is matching each tool to the job and to the data it needs to see.
Mistake five: no rule for what data the tools may see
Customer contracts, pricing exceptions and a prospect's internal documents end up pasted into whichever chat window happens to be open, under personal accounts nobody manages. Then a rep leaves and the context goes with them.
Decide in writing which data each tool may see, under which company account, and who owns that decision. It is a one-page document, and it is the difference between AI use the company can build on and AI use scattered across individual browser tabs.
The opportunities most teams miss
The mistakes are about doing old things faster. The opportunities are about things that used to be too expensive to do at all.
- Win/loss analysis on every deal, not a sample. A model can read every call transcript and lost-deal note from a quarter and show which objection kills deals. Most teams have never had that view, because it used to take an analyst a month.
- Account research that goes deep instead of wide. Rather than a thousand shallow personalised openers, an agent can build a real brief on fifty target accounts, covering stated plans, hiring and recent announcements, for a person to act on.
- Sales enablement that stays current. Battle cards and objection guides go stale within a quarter. A model working from your latest competitor pages and call notes can keep a draft up to date for a product marketer to approve.
- Localisation at launch, not a year later. Entering a Spanish- or Arabic-speaking market used to wait for a translation budget. A model makes a strong first draft cheap and a native reviewer makes it right. The opportunity is entering the market earlier, not skipping the reviewer.
- Internal tools built in days. Claude Code and similar coding agents let one operator build the small, specific tools that go-to-market teams used to wait two quarters for engineering to prioritise: a lead-scoring script, a pipeline-hygiene report, a pricing calculator.
Where to start
Pick one go-to-market motion that already works and has a clear outcome metric. Add one AI tool to one step of it, with a person approving every output for the first month. Measure the outcome, not the activity. If it improves, widen the step; if it does not, you have learned something cheaply.
What does not work is buying seats for everyone, writing a prompt library and waiting for the pipeline to change. The tools are good enough now that the model is no longer the bottleneck. The bottleneck is how clear the plan is that the model is asked to carry out.
The author advises ECLACTRA™ and partners with it to deliver fractional CMO leadership to clients, through TF Business Solutions.
