Buying an AI Tool

One agency, fourteen account managers, three meeting-notes tools. The playbook, filled in.

CASE STUDY

The firm

Northway Creative is invented. The tools, prices and terms are not.

Fifty people in London, Ontario. Marketing agency, mostly retainer clients. Microsoft 365 Business Standard on all fifty seats, HubSpot as the CRM. Fourteen account managers run about 120 client calls a week on Teams. After each call, the account manager is supposed to put notes and next steps into the HubSpot contact record. About half of them do, a day or two later, from memory. The other half do not, and the operations lead, Jess, finds out when a client asks why nothing happened.

The COO, Marcus, asked for "an AI notetaker." Three account managers had already tried three different free ones on their own accounts, which is how the firm ended up with client recordings in three places nobody controlled. That is the point at which a method is cheaper than a purchase.

The shape, filled in

Jess and three account managers wrote the shape in forty minutes, before opening a single pricing page. Line one took most of it.

Two lines did most of the work later. Line six, the systems it must sit inside, named both Teams and HubSpot. That is what turned the fit gate into a real test. Line four, the ceiling, was set at CAD 8,000 a year all-in, which Marcus said out loud only after Jess offered CAD 15,000 and let him correct it.

Gate 0 and the ladder

They already pay for an AI tool. Microsoft 365 Business Standard includes Copilot Chat. Jess pasted three Teams transcripts into it and asked for notes and next steps. The summaries were fine. The sentence she came back with: it cannot join the call, and it cannot put anything in HubSpot.

That sentence put the job on rung three of the ladder, a point tool for one job, with the paid Copilot add-on as the rung-one option to beat.

The options and the real prices

Four tools were shortlisted. One of each kind that could plausibly do the job, plus the one everyone assumed.

All four are under the CAD 8,000 ceiling. CAD at 1.37, before tax, seats only. Fathom's Business tier was priced rather than Team because the HubSpot field sync lives on Business. That is the tier trap in the deal criterion, caught before the invoice rather than after.

The gates

Six gates, pass or fail, with the vendor terms pages open.

ChatGPT Business was the account managers' favourite going in. It failed fit: the job is capturing the call, and a general assistant cannot join a call. Someone would have to paste a transcript in after every meeting, which is the step the firm was trying to remove. Out at the gate, no score, no argument in the room. That is what the gates are for.

The other three passed. Data and security took the longest. All three state they do not train on business inputs by default. All three support SSO. Copilot inherits the firm's existing Microsoft permissions, which the firm already manages; Fathom's SSO is on the Business tier; Fireflies passed on admin-managed accounts. Every one of those is a line from a terms page, not a landing page.

The job, scored

Three account managers ran all three tools for two weeks on their real calls, about fourteen each. They rated the fourteen job questions from the trial, not from the demos.

Fathom scored 21.4 of 25. Its notes and action items landed in the HubSpot record within minutes of the call, and the account managers used them as-is on most calls. Fireflies scored 20.0, close behind, with the HubSpot push needing a little more setup per person. Copilot scored 14.3. The recap in Teams was good, and it stayed in Teams. Getting it into HubSpot meant copying it, which is question five, and counting the steps, which is question six.

The rest, scored

The adoption, weight 4. Copilot 16.0. It was already in Teams, no bot joined the call, and nothing had to be set up. Fathom 14.0 and Fireflies 12.0: both put a visible recording bot in the client's meeting, and two clients asked what it was in the first week. That is a real adoption cost, and it is the kind that only shows up in a trial.

The trust, weight 4. Copilot 16.8: the firm's existing Microsoft permissions applied, and the terms were already in place. Fathom 14.0 and Fireflies 13.2: both fine on training and deletion; both required reading a new set of terms and deciding where client recordings would live.

The deal, weight 3. Fireflies 11.1, the cheapest of the three at fourteen seats. Copilot 9.9, an add-on on top of a suite they already pay for, annual commitment. Fathom 8.1, the most expensive because the HubSpot sync lives one tier up.

The scorecard

1.2 points between first and third. Anyone who tells you a scorecard makes the decision has not run one. The scorecard's job is to get you here: three tools that all do the job, and a decision that now depends on one or two things that were never on the list.

The tiebreaker

Three things, written on the scorecard page.

The bot in the client's call. Both point tools put a visible participant in every client meeting. Two clients asked. The firm decided that was manageable with one sentence in the meeting invite, but it was a real cost, and it favoured Copilot, which records nothing visible.

Jess already runs HubSpot. The tool that syncs to HubSpot has an owner on day one. The tool that lands in Teams needs a new process to get notes out of Teams, and nobody owned that.

Copilot's notes never left Teams. That was the whole job. The highest adoption and trust scores in the room belonged to a tool that did not do line one of the shape without a person typing.

What they did

Fathom Business, fourteen seats, monthly billing for the first quarter, then annual if week-three usage holds. One sentence added to every client meeting invite explaining the notetaker. Jess owns it. Marcus signed a one-paragraph decision the same day. The ninety-day review is in the calendar for December, with Copilot Business as the named fallback if clients push back on the bot.

Nothing about that decision was on the vendor comparison pages. All of it came out of the shape, the gates and a trial run by the people who do the job.

If you want to read the method first, the five articles are free at practicalmethods.ca/guides/buying-an-ai-tool.

The list is not the decision. It is how you find the one or two things that are.

Not a review, not rankings, not software advice. A method.