Revisiting a heuristic
Last spring I logged six weeks of using a language model in growth work and landed on a simple rule: it helps when I already know the answer and need it expressed, varied, or stress-tested; it costs me time when I don't know the answer and hope it does.
A year and considerably better tools later, that rule has held. What's changed is that the useful territory has expanded, and it's now clear enough to map properly rather than describe anecdotally.
The four places it has earned a permanent slot
1. Unstructured text at volume.
This is the largest genuine win and it's barely discussed, because it isn't glamorous.
You have 400 sales call transcripts, 1,200 support tickets, and three years of closed-lost reasons written as free text. Historically, nobody read them. There was no realistic way to — so this material sat in systems as a permanent, expensive blind spot.
Now you can classify all of it in an afternoon. Extract objection categories from every lost deal in 18 months. Cluster support tickets by underlying job-to-be-done. Find the language customers actually use to describe the problem, at scale, rather than from the six interviews you had time for.
This is qualitative research at quantitative scale, and it wasn't previously available at any price a Series A company could pay.
2. First-draft expression of a known argument.
Unchanged from last year. I have the analysis; I need prose. Saves an hour, produces something I rewrite.
3. Structured critique.
Also unchanged, and still the use I'd defend most strongly. Before a recommendation goes to a client: "what's the strongest case against this, and what would have to be true for it to be wrong?" It's a reliable check against my own pattern-matching.
4. Operational glue.
The boring one. Turning a spreadsheet into a formatted brief, converting research notes into a structured template, drafting the regex I can never remember, writing a first pass of a query. Individually trivial, collectively hours a week.
The three places it keeps being oversold
1. Strategy.
Ask for a channel strategy and you get a competent average of everything written about channel strategy. It reads well and contains nothing specific to your situation, because strategy is fundamentally about what you decline — and a model optimising for a helpful, complete answer will not tell you to stop doing four of the five things you asked about.
2. Content volume as a strategy.
The most common 2024 mistake I'm watching companies make. Production cost fell, so output went up tenfold — and traffic didn't follow, because the constraint was never production capacity.
If your content wasn't working before, it was because it lacked a specific point of view, original evidence, or genuine expertise. Producing ten times more of the same thing changes none of that, and it degrades the signal of the pieces that were working.
3. Personalisation at scale, done naively.
Generating a bespoke opening line for 2,000 cold emails by scraping LinkedIn produces messages that are technically personalised and obviously automated. Recipients have learned the pattern faster than senders expected, and the response rate collapse is measurable.
Where personalisation works is the top tier of an account list, where the research is real and a person reviews the output. That doesn't scale to 2,000, and pretending it does burns the channel.
The organisational question underneath
The teams getting real value have one thing in common, and it isn't tooling: they've decided what the model is allowed to touch unreviewed.
The failure mode I'm now seeing regularly is a team where AI-assisted output flows into client-facing or customer-facing material without a check, and quality drifts downward slowly enough that nobody catches it in any single instance.
My own rule: anything factual, numerical, or attributed gets verified by me. Anything expressive, structural, or exploratory doesn't need to be. It's a boundary worth writing down explicitly, because in the moment everything feels like it's in the second category.
What I'd do this quarter
If you do one thing: run the unstructured text analysis. Take every closed-lost reason and every sales call transcript from the last 18 months and classify them. Almost every team I've done this with has found at least one objection pattern nobody knew was there.
It's the highest-value application available right now, it uses data you already own and have never read, and it produces something no amount of content generation will.
Hilal Tasdan
B2B SaaS Growth Marketing Consultant & Fractional CMO. Partner in Growth.