Insight / signal
AI has moved the bottleneck upstream
AI is not just making delivery faster. It is exposing how slow the rest of the business is.
I keep seeing the same AI story with different company names attached.
A team builds something in two weeks that used to take a year.
A founder makes an internal business system in seven days that would previously have needed a proper team, a real budget, and a lot of meetings.
A software delivery company rolls AI agents across engineering, legal, project management, finance and commercial work, then realises the bottleneck is not just writing code anymore. It is requirements. Planning. Stakeholder coordination. All the bits that happen before and around the work.
That is the interesting part.
Not “AI makes people faster.” We know that now. Or at least we know it can, in the right hands, with enough taste and enough supervision.
The better point is this:
AI is moving the bottleneck upstream.
For years, agencies and software teams could hide behind production time. The site will take eight weeks. The campaign will take six. The tool needs a discovery phase, then design, then build, then review, then another review because someone important was on holiday and did not see the first one.
Some of that was real. Some of it was process theatre. Some of it was just the cost of human production.
AI is eating a chunk of that production drag.
OpenAI’s recent Wasmer case study says the team used Codex to build a Node.js runtime for the edge in two weeks. They say it would have taken a year without it. You can discount the exact number if you want — vendor case studies are not neutral documents. But even if the claim is half true, it still changes the maths.
The Endava case is more interesting for agencies and business owners. Their CTO says AI-assisted coding and agentic workflows meant the bottleneck was no longer engineering output. Requirements gathering, business analysis, planning and stakeholder coordination had to speed up as well.
That sentence is doing a lot of work.
Because this is where most companies are not ready.
They are buying AI as a faster production tool. Faster drafts. Faster code. Faster research. Faster reports. Faster social posts. Faster mock-ups. Faster internal apps.
Useful. Obviously.
But if the brief is vague, AI just helps you produce the wrong thing faster.
If nobody knows who can approve the work, AI just creates a bigger pile of assets waiting for someone to make a decision.
If the business has no decent source of truth, AI will rummage through stale context and sound confident anyway.
If the sales process is messy, AI will automate the mess.
If the marketing strategy is beige, AI will generate beige at scale. A crime against bandwidth, if nothing else.
This is why I think the agency conversation is slightly backwards.
A lot of people are asking whether AI kills agencies because clients can now generate copy, designs, reports and code themselves.
Some low-value agency work will get hammered. It should. Charging serious money for generic output was always a shaky business once the tools caught up.
But the better agencies, or whatever replaces them, will not be output shops. They will be operating-layer builders.
That sounds less glamorous. It is also where the money is.
When production becomes cheap, the valuable work shifts towards deciding what should be produced, why it matters, how it fits the business, who signs it off, what data it uses, how it gets measured, and what happens when it fails.
In plain English: the brief matters more, not less.
The approval system matters more.
The source data matters more.
The judgement matters more.
The annoying, unsexy workflow design matters more.
Take a campaign. AI can help with market research, positioning options, landing page drafts, ad variants, email sequences, call scripts, sales follow-ups and reporting.
Great.
But what is the offer? Who is the buyer? What is the proof? What claims are allowed? What does the sales team actually need? Which channels matter? Who approves the legal wording? What happens to leads after the campaign? How do we know whether it worked?
If those questions are fuzzy, the AI is just a very fast intern with access to a printing press.
Or take a client website. You can now generate page structures, metadata, schema, content variants, prototype components, analytics summaries and technical fixes much faster than before. But if nobody has decided what the site is supposed to do commercially, the speed does not help much. You get a prettier mess. Maybe a more modern mess. Still a mess.
Same with sales follow-up. AI can draft responses, score leads, summarise calls, suggest next actions and keep the CRM cleaner. Lovely. But if the business has not defined what a good lead looks like, when a human should step in, what tone is acceptable, which promises sales can make, and what happens after a hot reply, the system will wobble.
This is the part business owners need to hear.
AI does not remove operational discipline. It punishes the lack of it.
Slow production used to hide bad thinking. You had time to wander around the problem because everything took ages anyway. With AI in the loop, the work can come back before the decision is ready. That feels impressive for about five minutes. Then it becomes awkward.
“Here are 40 options.” Fine. Which one is right?
“Here is the prototype.” Fine. Who owns it?
“Here are the campaign assets.” Fine. What are we allowed to claim?
“Here is the report.” Fine. What decision does it change?
This is why the next serious AI services will not be sold as prompt packs or generic agent builds.
They will look more like operating systems for specific commercial loops.
A proper AI campaign operating system is not just a content generator. It should connect research, positioning, asset creation, approvals, publishing, sales follow-up, measurement and learning. It should have a source of truth. It should know which claims need approval. It should preserve evidence. It should make decisions easier, not bury the team in drafts.
An AI sales workflow is not just an email writer. It should know the lead stage, the offer, the previous conversation, the allowed next step, the handover point, and the risk of sending anything without a human in the loop.
An AI reporting workflow is not just a dashboard summary. It should say what changed, what probably caused it, what decision is needed, and what the team should stop doing.
That is the shift.
Not more output.
Better operating loops.
There is a warning here for business owners, but also a decent opportunity.
If you are currently slow because your team is overworked, AI can help.
If you are slow because nobody can make a decision, AI will expose you.
If you are slow because your process is full of handovers, unclear ownership and half-remembered Slack threads, AI will not magically fix that. It may make it worse by adding more movement to a system that already lacks control.
The first job is not to automate everything.
The first job is to find the repeated commercial work where speed would actually matter: turning a lead into a useful follow-up, turning a sales call into actions and assets, turning campaign results into the next test, turning a product update into messaging and distribution, turning a messy knowledge base into support answers, turning client notes into delivery tasks.
Then you design the loop.
Inputs. Sources. Owner. AI role. Human role. Approval point. Output. Measurement. Failure path.
Boring? Yes. Useful? Also yes.
This is where I think the post-agency model is heading.
Not “give us your brief and we will go away for six weeks.”
Not “we can make 500 pieces of content a month.” Please, no.
More like: we will take this repeated business process, make it AI-assisted, measurable and safe enough to use, then keep improving it as the evidence comes in.
That is a better offer. It is harder to fake. It is closer to the money. It is also much more defensible than being another content shop with a ChatGPT subscription and a Canva account.
AI has made production faster.
Now the brief is the brake.
The companies that deal with that will move properly faster. Not just busier. Faster.
The ones that do not will drown in options, drafts, prototypes and dashboards while still waiting for someone to decide what they actually want.
That is not an AI problem.
That is an operating problem.
And, frankly, it is a much better problem to solve.