Don’t go T-shaped.

Every software company starts as one good idea. A CRM that isn’t miserable. A database of prospects. A way to get twelve people onto a video call without filing a support ticket.

Then it draws a crossbar.

Attio started as a CRM and now ships automations and reporting. Apollo was a database of prospects and now sells sequencing, a dialer, meeting booking, call recording and a deal pipeline. Zoom won by being the one conferencing product that simply worked, and now sells docs, a whiteboard, mail, a calendar and an AI assistant to summarize the meeting you are already sitting in.

The shape has a name in the decks: T-shaped. One deep vertical stroke — the thing you are genuinely great at — and a wide horizontal one across the top, covering everything adjacent. A suite, but grown from a single product. A platform, but without buying anybody.

It is very close to being a good idea. It is also the most reliable way I know for a good company to become an average one.

The crossbar used to be expensive. A second product meant a second team, a second on-call rotation, a second designer who cared. Most companies could not afford one, which is why most companies stayed one product for a decade. That constraint is gone. What was two quarters for six people is a fortnight for one person and a fleet of agents. So everyone is drawing the crossbar at once — not because the strategy improved, but because the price fell.

What fell is the cost of building it. What did not fall, at all, is the cost of caring about it.

This is not a stupidity story. Every company drawing a crossbar has a spreadsheet saying it should, and for the first two features the spreadsheet is right.

Expansion revenue really is cheaper than acquisition. Selling a second product to someone who already trusts you costs a fraction of finding someone new, and net revenue retention is the number that decides your valuation. The data argument is real too: Apollo genuinely does know who you are prospecting, so sequencing built on that knows more than a standalone sequencer that has never met the contact. The defensive argument is realest of all — if you do not take the adjacent product, the adjacent product takes you. Apollo adds sequencing, Outreach adds data, and both of them are defending, not attacking.

And buyers do prefer fewer vendors. One invoice, one security review, one line in the renewal. Nobody’s job title is Head of Vendor Proliferation.

So the first adjacent feature is usually good. It is built by the same people, for the same users, close enough to the core that the team has real opinions about it. The second is fine. The failure is not at the second. The failure is in the tail — and the whole point of the crossbar is that it has a tail.

The 2026 version of the pitch is that the constraint which used to stop the crossbar has been removed. It has not been removed. It moved.

Agents are extraordinary at production. Give one a well-specified feature and it will write it, test it and document it faster than the team that specified it can read the result. What comes out is a first draft with the correct shape.

Everything after the first draft costs exactly what it always did.

Somebody still has to know that the fourth screen is where people get confused, because they watched them get confused. Somebody has to hold the reason the scheduler behaves strangely in Sydney, so that the obvious fix six months from now doesn’t quietly undo it. Somebody has to be embarrassed when it is bad. That is the entire difference between a feature and a product, and no model emits it, because it is not a thing you emit. It accumulates inside a person who is paying attention to one thing.

A team of eight people with a thousand agents still has eight units of attention. Spread across one product, every part of it has an owner who would notice. Spread across forty, each part has four percent of a person — and attention does not divide the way compute does. Below some threshold an owner is not a partial owner. They are an absentee.

Anyone who has run a suite knows how this actually plays out. The support ticket about feature thirty-one lands in the same queue as a bug in the core. It loses. It should lose. Which is precisely why feature thirty-one never gets good, and why it will still be not-good in two years, and why the person who files that ticket will eventually pay for something else to do that job.

Suppose you solve the attention problem. Suppose you are rich enough to give every feature a real owner and a real team. There is a second constraint underneath, and this one you cannot spend your way out of.

A suite is a shared ontology. One data model, one permissions system, one navigation grammar, one idea of what a record is. That coherence is the entire point — it is what makes a suite feel like a product rather than a folder of products. It is also a tax that the eleventh feature pays and its standalone competitor does not.

Google Meet is a good product built by good engineers. It also has to be a thing that lives on a calendar event, because that is what a meeting is inside Workspace: a row with attendees, a time and a link.

Almost everything people prefer about Zoom descends from that one noun being different. And Google cannot fix it by trying harder. Making a meeting a first-class object with its own lifecycle would fork it away from the thing that makes Workspace a suite in the first place. The constraint is not effort. It is that the crossbar must be expressible in the vertical’s vocabulary.

Every strong point solution wins on an opinionated noun. Linear’s is the issue and its graph of relations. Figma’s is a canvas with other people on it. Stripe’s is the payment intent — a payment modelled as a state machine rather than a transaction that either succeeded or failed. Those choices are why the products are good, and every one of them would have been vetoed inside a suite, because a suite already had a word for that.

So the crossbar is structurally condemned to ship the second-best version. Not because the team is worse. Because the good version needed a noun the suite is not allowed to have.

There is a third mechanism, and it is the one that finishes the job.

Software gets good through contact with people who are one click from leaving. That is the whole mechanism. A user who chose you will tell you, at volume and with feeling, exactly what is wrong — because they are invested and because they have somewhere else to go. A user who simply received you tells you nothing. They do not complain about the bundled whiteboard. They open it, sigh, and paste the result into the tool they actually like.

Nobody churns off Google Meet. So nobody tells Google what is wrong with Meet.

Bundling severs the loop that makes software good. It substitutes distribution for choice, and distribution buys usage — which looks identical to preference on a dashboard and is not remotely the same thing.

Teams is the honest case. By seats it beat Slack decisively, mostly by already being installed in a tenant that had already paid, and it beat it hard enough that regulators pushed Microsoft into selling Teams separately from Office. By preference it did not win — which is why Slack keeps getting picked in the places where the people using it get to pick. Both things are true at once. The bundle won the procurement and lost the argument.

Which gives you a real test for any crossbar feature: is this a place where someone chooses? If the buyer is procurement and the metric is seats, the crossbar is a fine business. If the user is a specialist who does this all day and could install an alternative before lunch, the crossbar is a rounding error that cost you a roadmap.

This is the same point as the factory note approached from the other side. The learning is the asset. Bundling does not merely fail to produce learning for feature thirty-one — it removes the signal that would have produced it.

The obvious rebuttal is history. Microsoft Office ate WordPerfect and Lotus 1-2-3, and neither of them lost on quality. Bundles have a long record of winning, and the people betting on the crossbar know the record.

Here is what is different, and it is not sentiment.

The suite’s moat was never the breadth of its features. It was the customer’s integration cost. In 1995, running the best word processor and the best spreadsheet from two different vendors meant two file formats that would not talk, two support contracts, two upgrade cycles, and a macro language that stopped dead at the boundary. The bundle’s value was that it solved a problem the customer could not solve themselves. You bought the suite to avoid the seams.

That cost has been falling for twenty years — APIs, single sign-on, webhooks, every iPaaS ever funded — and in the last two it has fallen off a cliff. An agent that reads and writes across six tools is a better integration layer than any suite’s internal consistency, because it integrates the six tools you chose rather than the sixty somebody shipped.

Which produces the asymmetry I think decides this whole question. The suite has to build and maintain the union of everything anyone might want. The customer only has to assemble the intersection of what they actually use. AI lowered the cost of both jobs. It lowered the customer’s far more.

None of this says bundles are dead. It says the winning bundle is not grown. It is assembled.

Salesforce did not build Slack, or Tableau, or MuleSoft. It bought them — each one after it had already won its own category, against real competition, with its own users who had chosen it on purpose. What Salesforce assembled was a set of products that were each somebody’s favourite thing.

That is not a T. It is a comb: several full-depth strokes held by a thin spine. The teeth stay deep because each kept what made it deep — a team that owns it, users who chose it, and its own nouns.

Bundling by acquisition preserves the exact thing that bundling by extension destroys.

And here is the part that should worry anyone drawing a crossbar this quarter: the customer is now assembling a comb too. Six best-in-class tools, an agent layer running across them, and no patience for the seventh product that arrived free with the third. They never needed your crossbar. They needed your vertical to be the best one they could find.

Watch the same thing happen to firms rather than products.

A three-person consultancy that spent nine years becoming the best in the country at one narrow thing now discovers it can plausibly offer twelve services. The agents write the deck. The agents do the research. The marginal cost of a new service line went to roughly nothing, so why not take it?

Because your pricing power was never your capacity. It was that nobody else runs your method. The moment you offer everything, you are a generalist competing with every other generalist whose agents also got cheap — and generalists get priced by the hour, because there is no reason on earth to pay a premium for the second-best option at anything.

The firms that get destroyed by cheap production will not be destroyed by AI doing their work. They will be destroyed by doing what the software companies are doing: reading the collapse in production cost as permission to widen when it was a mandate to deepen.

That last one is the whole trap. What is cheap to build is cheap to copy. Anything an agent produces for you in a fortnight, your competitor’s agent produces for them in a fortnight, and the arms race ends with four suites carrying the same eighty mediocre features and exactly one real difference between them: the stroke each of them started with.

So spend the collapse in production cost on depth instead. Same headcount, same agents, all of it aimed at the thing you are already best at, until the gap is embarrassing to look at. Depth is the only output that does not get commoditized by the tool that produced it, because it is made of judgment about one domain and that is the one input nobody can generate.

We are trying to hold this line at Ergoly, and it is harder than it sounds. We build one thing: workers that carry a specific firm’s method — how this firm scopes, prices, argues and decides — and keep that method inside the firm. The depth in it does not come from our roadmap, which is the only reason a company this small gets to be deep across ten thousand different domains at once. It is supplied by the firm. Our job is to be excellent at holding it and to keep our hands off the crossbar.