10 things to get right with AI in 2027
We spent most of this year building agents, for clients and for ourselves. Here's what we learned, and what I'd get right going into 2027. If you only have a minute, this is the short version.
The short version
Your next customers will increasingly arrive through agents, so your website has to be loved by humans and preferred by agents, and your brand needs real craft to stand out from the slop. Inside the business, the value comes from owning your data layer and testing agents properly before you trust them. Measure use, share and trust rather than tokens, so you can see which work actually got better.
None of it sticks without people. Train them while you build and turn one person's best AI tool into everyone's, but keep them in the work that needs judgement. Once the agents are running, give them a chief of staff so they save your time rather than eat it, and keep tuning, because AI is a cycle with no finish line. The detail behind each is below.
The detail
1. The next billion customers are agents
Personal agents are starting to shop and book on people's behalf. Meta launched Muse in September, a month after xAI launched Grok Bot in August. As the next billion customers come online through agents, your website has to be loved by humans and preferred by agents. B2B is already moving: G2 found 51% of software buyers now start their research with an AI chatbot more often than with Google. Since we started taking our AI visibility seriously and building an "agent mode" on the site, inbound leads have grown 4X in three months.
2. Measure what matters, not tokens
Several leadership teams have proudly shown us their AI reporting, and it turned out to be a token leaderboard: who spent the most this month. That's a trophy for the biggest bill. When we audited one client's top users, they weren't doing the most valuable work. They'd left the most expensive model running on tasks like finding an email. We measure in three stages: use (are people showing up), share (is value spreading across teams) and trust (how much real work is handed to AI). McKinsey's latest survey found only 37% of companies see any impact from AI on profit, and the few who see a lot are twice as likely to have a defined way of measuring it. If your AI reporting disappeared tomorrow, you should still be able to say which processes got faster or cheaper.
3. Earn trust before you hand over the keys
An agent is only useful once people trust its answers, and that trust has to be tested rather than assumed. When we built Tesco's creative compliance tool, the models out of the box were only right 50% of the time on whether an ad was compliant. We built an eval suite from hundreds of Tesco's real approval rules and kept tuning until confidence passed 90%, which is when the agent was good enough to run with less intervention. Campaign creative that took up to four weeks to go live now goes live in days. Most DIY agents skip this step, which is why nobody can say whether they can be relied on.
4. Bring your brand story to life
Paired with design, AI is the best storytelling kit we've ever had. A vision used to live in a strategy doc or a deck, and people had a hard time understanding it. Now you can show it as a working prototype in days, then use AI motion to turn it into a short video with a customer at the heart of the story. And as AI fills the web with slopsites that all look the same, craft is what makes a brand worth remembering. That needs to live on as a strong design system: shared components and rules that your team and your agents build from, so every prototype and deck comes out looking distinctly you. Our own agents now build new pages for our website and slides for our decks inside our brand system, generating particle visuals from our brand and pulling case study assets from hundreds of past projects.
5. Own your data layer
We're seeing more businesses who want to centralise and own their data. An AI assistant inside one tool only sees what's in that tool's walled garden, so it never gets the full picture of your business, clients and projects. Build your own data layer above the models, and the use cases get much richer, because agents can work across everything your business knows, and switching models never means starting again. Keep access and action separate. Letting agents read from it is low risk. Letting them write back to it, or act on what they find, needs tighter permissions and an owner, with a clear rule for what gets promoted from draft to fact, or it fills up with slop.
6. Spread your best use-cases
The best AI use cases in most businesses are sitting in one person's chat window. Someone builds something brilliant, nobody else knows it exists, and when they leave it leaves with them. Our own sales coach started as one person's routine: it preps each call and drafts the follow-up afterwards. We promoted it to a shared skill, and now everyone with a commercial part to their role uses it. That runs through our shared skills hub. Every skill is versioned, and a submission flow checks it and gives the builder feedback before it goes for central approval. Role-based groups then decide who can use which skills. When the team said its emails sounded off, we fixed the voice once and every agent got better. Use regular feedback & interview rhythms to understand what's working and what isn't, so you can promote winning skills company-wide and tune the areas that aren't working well.
7. Train your people while you build
Building the systems is half the job. While that happens, people still need to learn how to use assistants & agents in their day-to-day work, and how you want them to work alongside AI. The setup we've seen work best starts with a named AI lead, or a fractional Chief AI Officer, who owns the strategy & rollout. Then add an AI champion in each discipline and bring them together as one AI excellence working group, so what works in each team can benefit the wider company. Managers have to use it themselves too. Lewis, our behavioural strategist, calls it the messenger effect: who uses a tool tells everyone else whether it matters. Regular office hours with the AI lead means that the business can quickly unblock work with a central steer.
8. Keep people in the work that needs judgement
The more we trust AI, the less we may think. This points to the IKEA effect: we value what we've helped build, and we value an output less when all we did was approve it. So delegate the work nobody needs to learn from, and where judgement is the point, design the agent as a coach. On a go-to-market agent we built for a large enterprise, the agent questions the seller and pushes back on their thinking, while a score rises as the pitch gets stronger. Sellers put more in because they can see it building.
9. Give your agents a chief of staff
Here's the problem nobody warns you about. Once you have fifteen agents working for you, your inbox and your day get busier, not quieter, because every agent wants an approval or a decision. The fix is a reporting line: one chief of staff agent that the others report into, which filters the noise and bundles requests into a few set moments in the day. Anything urgent still gets through straight away. A person should always be accountable for what an agent does, so this scales as a network rather than a single hub: everyone who owns agents gets their own chief of staff, and each agent reports to the person accountable for its work. Our sales agents roll up to me, and our engineering agents to our CTO. We're building this for ourselves right now, and it's the difference between agents saving your time and agents eating it.
10. AI is a fierce cycle
What's possible this month wasn't possible last month. Models change, and so do your processes, your data and what people expect. Nothing you build stays finished, and an agent that was sharp in Q2 can quietly fall behind the next quarter. AI isn't a project with an end date. It needs a constant cycle of tuning and trying new things, with someone whose job it is to run that cycle. Tara Seshan, who leads ChatGPT Work at OpenAI, says she builds her roadmap for where the models will be in two to three months, not where they are today. That's why, for a lot of our customers, this now runs as a managed service rather than a one-off project. It has three parts: a data and systems strategy, rapid cycles of prioritising, building and tuning, and the people and culture work that drives adoption. None of it is a one-and-done job. The roadmap gets revisited every quarter, because by then the models will have moved again.