I write about AI a fair bit, mostly favourably, mostly about specific tasks where it earns its keep. So I decided to push further: for one working week, every decision in my sales pipeline that AI could make, AI would make. I’d review, but I wouldn’t override unless something was actively wrong.

Five days. Twenty-three live opportunities. Here’s what actually happened.

The setup

I run a small consultancy. My pipeline at the start of the week had:

8 new leads from the previous fortnight
11 quoted opportunities awaiting a decision
4 won deals waiting on signed contracts

For the experiment, I gave AI access to:

Lead source and conversation history
Quote details and dates sent
The pipeline stages and how long each was in them

And asked it to:

Suggest the next action for every opportunity each morning
Draft any follow-up emails
Flag anything stuck or going cold
Score leads by likely-to-close

I’d skim the suggestions over coffee, approve or modify, and let it execute.

Day one: surprisingly good

Morning one, I expected to spend an hour fixing AI’s bad suggestions. I spent fifteen minutes.

The lead scoring was sensible. The two leads it flagged as “unlikely to close” were both ones I’d been quietly suspicious of. The follow-up drafts were polite, specific, and included references to the actual quote details — not generic boilerplate.

One thing it got right that I wouldn’t have: it noticed a quote sent 11 days ago to a returning client and prioritised the follow-up. I’d have left that one another week, because I’d subconsciously thought “they always come back.” AI didn’t have that bias. It saw stale and acted on it.

Day two: the first miss

A new lead came in via the website form. AI’s suggested response was technically correct but tonally off — too formal for the actual enquiry, which was a casual question from someone who’d seen me speak at an event.

I overrode it. Sent a much shorter, friendlier reply. Got an enthusiastic response within an hour.

This was the first sign of what would become a pattern: AI is good at the standard response but it doesn’t read social context. If someone signs an email “cheers, mate,” they don’t want a reply that opens “I hope this email finds you well.”

Day three: it caught something I missed

A client I’d quoted three weeks ago had gone quiet. AI flagged it as “high priority follow-up” because the quote was about to expire and the deal value was significant.

I’d genuinely forgotten. Not “deprioritised it” — actually lost track. AI didn’t lose track because AI doesn’t have a brain that gets distracted by other things.

The follow-up went out. The client replied apologising for the silence and asking to extend. We’re now in active negotiation on a deal that would have died.

That single intervention covered the cost of the entire experiment ten times over.

Day four: the limits start showing

A long-time client emailed asking about a job. AI’s draft response treated the enquiry like a fresh lead — explaining our process, attaching a brochure, the works.

This was a person I’d worked with for four years. We’d had a beer together. The right reply was three lines: “Yes, can do. Can we chat Tuesday?”

AI doesn’t know which clients are old friends. It can’t, unless you’ve meticulously tagged them, and even then it doesn’t feel the difference. The “treat everyone identically” mode that makes AI consistent is also what makes it occasionally tone-deaf.

Day five: trust calibrates

By Friday I’d settled into a rhythm. AI handled:

Initial responses to new leads (after I checked the tone)
All quote follow-ups
Stage progression suggestions
Stale-deal flags

I handled:

Anything involving an existing relationship
Anything sensitive (price negotiations, awkward conversations)
Final approvals on quotes
The actual deals where strategy mattered

Net effect: my morning admin time dropped from ~90 minutes to ~25. Two opportunities were rescued from quiet death. One lead got an off-key response that I caught before it sent.

Should you do this?

Honestly, yes — but with the framing right. AI isn’t going to “run” your pipeline in the sense of replacing your judgement. It’s going to handle the routine attention that humans are bad at, and free you up for the work that actually requires you.

The way to think about it: AI is the ultra-reliable junior employee who never forgets, never gets distracted, and never has a bad day. It also has no feel for nuance. The trick is matching the work to the worker.

stedd.io has these AI features built into the pipeline, so you can run a similar experiment without the setup cost.