Agents aren’t colleagues: why 2026 is the year AI becomes practical (as workflow automation)

In New York in December, the most interesting AI conversations weren’t about models giving smarter answers. They were about systems that can do the work.

(Yes, I'm having a bit of fun with the tiger image here. Which is also the point: powerful tools are arriving fast, but they only become useful when we rein them in)

Not in a sci-fi, “AI colleague” way. In a much more useful way: taking messy, repeatable workflows, the kind every business runs on, and making them faster, cheaper, and easier to scale, with the right checks in place.

Since then, one thing has become even clearer: AI has a brand problem.

The story sounds like bold autonomy. The reality that works inside organisations is more grounded, and far more actionable.

AI’s brand problem: “agents as colleagues” vs what organisations can actually deploy

If you listen to the narrative, “agentic AI” sounds like autonomous digital staff.

If you look at what’s working in practice, it’s closer to this:

Agentic AI, in practice: workflow automation upgraded with language + tool access + orchestration.

That framing matters because it shifts the conversation from fear to design:

• What’s the workflow?

• What can be automated safely?

• Where do humans sign off?

• What gets logged?

• What outcome are we measuring?

This is how I use agents with clients: keep humans accountable, automate repeatable steps, and design the guardrails up front. No “colleagues”. No theatre. Just better systems.

Why 2026 looks like an inflection point with AI workflow automation (even if most firms aren’t “there” yet)

My base case for 2026 is not that everyone suddenly “wins with agents”. It’s that the platform layer makes them hard to ignore, and the economics make them hard to postpone.

Gartner’s headline prediction is blunt: by 2026, 40% of enterprise applications will include task-specific AI agents, up from less than 5% in 2025.

At the same time, executives are moving from experimentation to intent. Bain’s executive survey from late last year described AI shifting from pilots toward production, with leading use cases (including software development and customer service) already scaling, and AI becoming a top strategic priority for many leaders.

Put those together and you get the practical version of “agentic”:

2026 is when more businesses stop asking “should we use AI?” and start asking “which workflows should we redesign first?”

Design is where this is becoming real fast (marketing benefits, but it’s bigger than marketing)

The reason I’m increasingly excited about 2026 is that we’re seeing a step-change in something that matters across many functions:

turning ideas into high-fidelity visuals quickly, with increasing control.

Google’s Nano Banana Pro is one example of where the tooling is heading. Google positions it as an image generation and editing model aimed at “studio-quality” precision and control, including better text rendering (useful for posters and diagrams), localisation, and professional-grade creative controls.

Marketing teams notice this early because they live in a world of variants and asset volume. But the deeper shift is broader:

1) Architects and designers: faster iteration, earlier alignment

Instead of waiting for a perfect render pipeline or a full prototype, teams can create credible visuals earlier — not as final truth, but as a tool for alignment:

• explore more options quickly

• converge on a direction sooner

• reduce the time between “concept” and “something you can react to”

2) Product teams: prototypes before prototypes

For many product teams, the real drag isn’t creativity — it’s the calendar time between concept → mock → proto → iterate. High-control visual generation and editing compresses that loop, and it changes how quickly teams can align internally before they commit time and money.

3) Sales teams: moving beyond theoretical conversations

This is the under-discussed use case.

When clients can see what they want, quickly, conversations change. It’s no longer abstract requirements and vague preferences. It becomes: “this, but…”, “closer to that…”, “what if we combine…”. That accelerates discovery, reduces misunderstandings, and makes proposals feel tangible.

4) Marketing: the workflow becomes the system

Yes, marketing is still a prime playground, especially once tools get better at controlled variation (brand, product truth, local language, format).

But the real point is this:

Design capability becomes a workflow advantage across the whole business.

None of this is foolproof today. You still need review loops — for product truth, brand integrity, and basic quality control. But the direction of travel is obvious, and it increasingly feels like months, not years, before the friction drops again.

Yes, I’m having a bit of fun with the tiger here. It’s also the point: powerful tools are arriving fast, but they only become useful when you rein them in.

So… will 2026 be the last year of photo shoots?

No quite, but 2026 will be the year photo shoots stop being the default for a large share of content production.

What changes is the role shoots play.

Photo shoots shift from “the production method” to “truth capture”

In many categories, the winning pattern will be:

• capture a smaller set of high-quality “truth” assets (materials, fit, hero references)

• then use controlled generation/editing to multiply into variants

In other words: shoot less, reuse more, iterate faster.

AI replaces the long tail first, not the hero

AI will displace:

• variant-heavy ecommerce needs

• seasonal refreshes

• localisation

• performance creative testing

• supporting PDP assets

Shoots still win for:

• hero campaigns where trust is the point

• anything high-scrutiny (claims, materials, fit)

• tactile categories where authenticity is part of the brand promise

So the question isn’t “shoots or AI?” It’s “where is the truth required, and where is controlled variation the smarter system?”

The controls that make agents (and creative automation) safe enough to scale

If you want speed without chaos, treat controls as part of the product.

That’s not a compliance afterthought. It’s the difference between a workflow you can scale and a demo you can’t.

A practical checklist that holds up in the real world:

1. Define the workflow boundary

   What inputs does the agent/tool get? What outputs is it allowed to produce? What tools can it call?

2. Least-privilege permissions

   If it doesn’t need access, it doesn’t get access.

3. Human sign-offs at decision points

   Especially anything customer-facing, regulated, or brand-critical.

4. Audit trail

   Keep a log of what was generated/changed and why.

5. Measure outcomes, not novelty

   Cycle time, cost per asset, error rates, conversion lift, team capacity freed — not just “usage”.

This is also why the “AI colleague” narrative is unhelpful. Colleagues don’t come with permissioning models, logs, and approval gates. Systems do.

What this looks like in practice (patterns you can recognise)

Below are three patterns that translate well across organisations. They’re deliberately anonymised, but you’ll recognise the shapes:

Pattern A: Brief → variants → QA → approve (creative ops)

• Before: a human bottleneck for every variant; supplier time spent on resizing/reformatting rather than ideas

• Automation role: generate structured variants from a constrained brief + reference pack; route for QA; escalate exceptions

• Controls: brand constraints + product truth checks + mandatory human approval before publishing

• Outcome: faster iteration cycles; more tests per month; creative leaders spend time on concepts, not production

Pattern B: Always-on reporting that writes itself (performance cadence)

• Before: recurring weekly decks, manual pulling, inconsistent commentary

• Automation role: extract metrics, reconcile, draft narrative, flag anomalies

• Controls: source validation + confidence flags + analyst sign-off

• Outcome: leadership gets a consistent operating cadence; analysts focus on insights and actions

Pattern C: Supplier delivery with an “agentic” option (partner operations)

• Before: work priced and planned as if headcount is the only scaling lever

• Automation role: supplier proposes an automated workflow that removes manual steps and shortens cycle time

• Controls: defined SLAs + logging + approval gates + clear accountability

• Outcome: faster delivery at lower cost, with less dependency on hiring

The simplest way to start in 2026: ask suppliers for an agent-based option

If you do one practical thing this quarter, do this:

Ask every supplier (regardless of category) to include an agent-based workflow option in their proposal.

Not “do you use AI?”

But:

• where in your delivery workflow could automation remove cycle time?

• what would be automated, and what stays human-owned?

• what controls sit around it?

• what outcome metric will you commit to?

This shifts the conversation from buzzwords to operations, and it quickly reveals which partners understand what 2026 demands.

Closing thought: 2026 isn’t about AI colleagues, it’s about redesigning work

My working thesis for 2026 is simple:

• agents aren’t colleagues

• they’re not strategy

• they’re a way to implement process improvement faster, if you design them with accountability and controls

And design-led workflows, from architecture to product to sales to marketing — are where that shift will feel most tangible, most quickly.

Question I’m holding for the year ahead: Which workflows in your organisation will you stop scaling with headcount, and start scaling with guardrailed automation?