If your team already uses AI but marketing results have not changed, do not begin with another license or a department-wide transformation. Choose a recurring workflow with visible business friction, document how it performs today, and commission a bounded implementation that can operate within 90 days. Compare the new flow against a real baseline across time, quality, cost, and commercial outcome. The end product is evidence to scale, revise, or stop.
A useful first workflow could be lead handoff, customer reactivation, production of a recurring campaign, or a weekly synthesis of media signals for a decision. The right choice depends on your constraint and data. AI should handle specific tasks such as interpreting unstructured inputs, drafting controlled variations, or recommending a next step. Rules, integrations, and accountable people still carry much of the workflow.
Why access to AI rarely changes the outcome by itself
Individual tools can improve isolated tasks: summarize a document, draft a version, or generate ideas. A commercial outcome crosses briefs, data, approvals, channels, service, and measurement. If AI output returns to a slow queue or requires manual reconstruction, the team produces more drafts without shipping better work. The issue may not be adoption; it may be the operating design.
A Quarterly Journal of Economics study followed 5,172 support agents and found a 15% average increase in issues resolved per hour, with different effects by worker experience. That is evidence that impact depends on the task and user, not a marketing promise. Look for frequent, bounded work with an observable standard, enough repetitions to learn, and consequences your team can review.
Choose the first workflow through five decisions
- Outcome: the decision or deliverable that must improve and who consumes it.
- Repetition: the workflow occurs often enough to produce evidence during the pilot.
- Baseline: time, quality, cost, or conversion can be observed before the change.
- Reversibility: people can review output, handle exceptions, and return to the current process.
- Accountability: one owner covers the full workflow, including boundaries among the agency, marketing, sales, data, and IT.
Google PAIR recommends mapping the current workflow and checking whether AI adds unique value; a rule or heuristic may be easier to explain, debug, and maintain. Design the process before selecting a model. Fix unclear approval ownership as governance. Fix missing events as integration. Reserve AI for uncertainty or volume that warrants it.
The one-page brief you need before requesting proposals
- Authorized input: the data, documents, and signals that enter, plus explicit exclusions.
- Current work: steps, systems, owners, wait time, rework, and common exceptions.
- AI role: one task, expected output, autonomy level, and evidence attached to each response.
- Human control: who reviews, when they can pause, and which decisions remain human.
- Measurement: baseline, comparator, primary metric, guardrails, and observation window.
- Delivery: integrations, training, documentation, monitoring, and post-pilot operating owner.
This is not a maturity score. It is a scope agreement that lets a buyer compare an agency, a technology consultancy, or a joint delivery model. A proposal should separate process discovery, workflow design, configuration, integration, evaluation, change support, and operations. A model demo is not a substitute for any of them.
Days 1–15: map and measure the current workflow
Choose one market, channel, or demand type. Inspect complete cases, including failures. Record active work and wait time, volume, rework, approval, variable cost, and downstream result. Interviews explain the records, but the baseline needs comparable operational evidence. Do not count an automated acknowledgement as a handled opportunity or a draft as an approved asset.
Days 16–30: design the new division of work
Specify what remains deterministic, what may use AI, and where a person decides. Build accepted and rejected examples, quality criteria, and difficult cases. Use anonymized or synthetic data until real access is authorized. NIST treats measurement and evaluation as lifecycle work, so logging, prompt versions, source data, and a route to contest an output belong in the design rather than a later compliance pass.
Days 31–60: run in shadow mode
The system prepares an output while the current workflow retains control. Compare usefulness, accuracy, correction time, and failures by segment. Test incomplete inputs, stale content, multiple languages, hostile instructions, integration downtime, and priority changes. Shadow mode does not establish ROI. It shows whether the flow is safe and operable enough for a live comparison.
The UK government's People Factor guidance argues that value requires embedding a tool in routines and measuring use and impact, not simply giving access to a small set of enthusiasts. Its experience comes from a public-sector rollout and does not validate a private marketing project. The transferable lesson is narrower: training, support, and work design are part of the implementation.
Days 61–90: controlled pilot and operating decision
Release the workflow to a bounded cohort with an exception queue, monitoring, and a rollback path. When volume permits, compare equivalent cases with the current process. For productivity, measure time to accepted output, human hours, rework, and total cost. For revenue, keep a persistent comparator and respect the buying cycle. Add error, complaint, cancellation, margin, and team-load guardrails.
At the decision gate, ask more than whether the model worked. Did the full workflow create enough value? Can the team operate it? Which costs scale with volume? Which dependencies remain? Scaling may mean expanding one proven flow rather than opening five use cases. An inconclusive result should produce a documented next hypothesis, a longer window, better data, or a simpler non-AI solution.
Cost drivers and delivery responsibilities
Major cost drivers are process discovery, data access and cleanup, integration count, channels, model volume, test-set creation, human review, observability, support, and work redesign. Licenses are one line item. Ask for separate pricing to build, operate, and improve the workflow for three to six months. A low-cost proof can conceal the cost of reliable operation.
Marketing owns the outcome and quality standard. Data and IT own access, identity, integration, and reliability. An agency may redesign the journey, content, and campaign operation; a technical consultancy may implement integration and controls. With two partners, appoint one accountable end-to-end owner. The provider should not be the only party defining the metric that proves its own work.
Start in one market before regional rollout
US midsize teams often already have overlapping SaaS, agencies, and platform AI. Prefer a first flow that reuses the current stack and makes ownership visible. If the company operates across countries, currency, offer, channel rules, language, calendar, data access, and local ownership are requirements, not a translation phase. Validate one operating model before adding regional complexity.
Bring a real workflow to the first conversation
Use the broader prioritization framework at https://makinai.co/insights/en/practical-framework-prioritize-ai-use-cases and test whether the proposal needs AI at https://makinai.co/insights/en/evaluate-whether-ai-consulting-proposal-actually-needs-ai. See concrete workflow examples for lead handoff at https://makinai.co/insights/en/ai-lead-qualification-marketing-sales-handoff and reactivation at https://makinai.co/insights/en/ai-customer-reactivation-crm-without-discounts. MAKINAI's strategy and transformation service is at https://makinai.co/services/en/ai-strategy-transformation-consulting. Bring one workflow, anonymized samples, and the metric that must change; we can discuss which step is worth testing first.
The 90-day sequence is an editorial structure for a bounded first workflow, not an implementation guarantee. Complex integrations, legal review, limited volume, or long buying cycles may require a different window. The sources below support workflow selection, adoption, and evaluation principles; they do not promise a commercial result for this project.