Building the 2026 AI Roadmap — A Practitioner Framework

It's the time of year when enterprise teams build their annual AI roadmaps for the coming year. The exercise has matured significantly since 2023, when the planning was largely speculative. By late 2025, the planning has substance — existing workloads to operate, new workloads to deliver, capability to build, risks to manage.

This is a practitioner framework for building the 2026 AI roadmap — what to include, what to leave out, and how to produce a plan that survives the first quarter intact.

The components of a useful roadmap

A roadmap that drives action covers:

The current state

What's actually running in production. Specific workloads, specific user populations, specific value metrics. This is the baseline against which 2026 progress is measured.

Operating commitments

What the team will keep operating in 2026. Sustaining work — keeping systems running, responding to incidents, addressing technical debt, supporting users. Often 30-50% of total capacity.

Pipeline commitments

New workloads the team will deliver. Each with a clear sponsor, an articulated value case, a sized estimate, a target go-live. Not aspirations; commitments.

Capability investments

Investments in the underlying capability — platform improvements, tooling, evaluation infrastructure, governance enhancements. The foundations that enable workload delivery.

Risk and contingency

Known risks with mitigations. Reserve capacity for the unexpected. The plan that has no reserve is the plan that doesn't survive surprises.

Governance milestones

When new policies will be issued, when reviews will happen, when regulator engagements are scheduled. The governance work has its own timeline.

Measurement

How success will be measured. Specific metrics, with targets, with cadence. Without measurement, the plan can't be evaluated.

What to leave out

A few categories that tend to bloat roadmaps without value:

Capabilities that may emerge

The roadmap shouldn't depend on capabilities not yet available. Plan for what exists; allow for what might emerge but don't commit to it.

Workloads without sponsors

A workload listed without a clear business sponsor is a wish, not a plan. Sponsors evaluate; some wishes don't survive.

Aspirational metrics

Targets that aren't realistic given current trajectory. Better to commit to achievable improvements than to publish aspirational targets that won't be hit.

Buzzword projects

Workloads positioned around an emerging buzzword without a specific value case. The buzzword will be different next year; the value case must stand independently.

Vendor-driven priorities

Workloads on the roadmap because a vendor pitched them. The vendor's interest isn't the same as your organisation's interest. Filter accordingly.

The shape of a 2026 plan

Specific patterns we are seeing in 2026 plans being built:

More operating, less greenfield

Enterprises with significant existing AI deployments are spending more roadmap on operating and improving than on new workloads. The shift is appropriate; the existing investments need to compound rather than being abandoned for the next exciting thing.

Migration of early deployments

Many 2026 plans include migrating early-stage deployments to mature architecture — the work covered in our earlier piece on this category. The technical debt of 2023-2024 deployments is the substance of 2026 work.

Agent infrastructure investment

For organisations expanding into agent workloads, the platform investment is significant. Agent runtime, tool catalogues, evaluation, observability.

Governance maturation

Following 2025's codification of governance frameworks, 2026 plans include encoding more of the framework in the platform. Less manual review; more automated enforcement.

Specific workload areas

The dominant workload categories for 2026 plans we are seeing:

• Regulatory and compliance workflows (continuing the dominant pattern)
• Customer service augmentation
• Document workflow automation
• Code and engineering productivity
• Knowledge management
• Specific industry workflows (life sciences submissions, banking compliance, energy inspections, government services)

Multimodal expansion

Workloads adopting multimodal capability. Document processing with complex layouts; field operations with photographic evidence; mixed-media customer interactions.

Cost optimisation as a stream

Explicit work on cost optimisation. Migration to cheaper models where adequate; caching infrastructure; usage right-sizing. The savings fund other investments.

What's not in most 2026 plans

Some categories conspicuously absent or modest:

Autonomous agents at scale

Despite continued marketing, most enterprise plans for 2026 maintain human-in-the-loop. Autonomous agent ambitions are research bets, not production commitments.

Foundation model training

Few enterprises are planning to train foundation models. The economics and the capability requirements don't justify it for most. Fine-tuning of open models for narrow tasks is more common.

Generic AI strategy

Strategy documents at the abstract level have been largely replaced by specific workload roadmaps. The era of "AI strategy" as a deliverable is ending.

Quantum-AI hybrids and other speculative

Speculative capabilities aren't on serious 2026 plans. They're on research roadmaps, perhaps; not on operating roadmaps.

The planning process

A working process for building the roadmap:

Q4 — gather inputs

Workshops with business sponsors, technical teams, governance partners. What worked in 2025; what didn't; what's needed; what's risky.

Q4 — synthesis

Draft roadmap with sized estimates, risk assessments, governance milestones. Distinct from a wish list.

Q4 — alignment

Review with executive sponsors, governance partners, dependent teams. Adjust based on feedback. Commit to what you can actually deliver.

Q4 — publish

Document the plan with measurement framework. Each commitment with success criteria.

Quarterly — reforecast

Plans evolve. New information arrives; priorities shift; capacity changes. Quarterly reforecast keeps the plan current.

End of year — retrospective

What we said; what we delivered; what we learned. Inputs for the next year's plan.

This is conventional programme management applied to AI. The discipline is the same as for any other enterprise programme.

What we keep seeing

Patterns in 2026 planning cycles we have been engaged with:

Plans are more grounded than they were. Less speculation; more commitment. The era of breathless AI strategy is replaced by structured workload roadmaps.

Operating dominates new delivery. Mature deployments need sustained investment. New workloads compete with operating commitments.

Governance work is significant. Regulatory expectations are crystallising; institutional capability needs to keep pace.

Cost discipline is mainstreaming. Cost optimisation is a planned stream of work, not an afterthought.

Cross-functional planning is the norm. AI roadmaps now include risk, compliance, legal, business inputs as primary, not as review-only.

Vendor portfolios are being rationalised. Multiple vendors accumulated through 2023-2024 are being consolidated where it makes sense.

What we recommend

For enterprise teams building 2026 AI roadmaps in late 2025:

1. Start with what's currently running. The 2026 plan is built on top of 2025's reality.
2. Allocate generously to operating and improving. New workloads need foundation work.
3. Commit only to what has a sponsor and a value case. Wishes belong elsewhere.
4. Build in governance milestones explicitly. The work has its own timeline.
5. Reserve capacity for the unexpected. Plans without reserve don't survive.
6. Plan quarterly reforecasting. The annual plan won't be right; the discipline of adjusting it will.
7. Engage business sponsors as partners, not as reviewers.

A 2026 AI roadmap that survives is built deliberately. The roadmaps that don't survive are built once and put on a shelf. The discipline of planning, committing, executing, measuring, and adjusting is what distinguishes the enterprises that deliver AI value from the ones that produce activity without value. The 2026 plans are now being written; the discipline they reflect will shape next year.