AI & Agentic Workflow Engineering
Put AI to work in your marketing and revenue operations.
AI creates opportunities to rethink how teams research, make decisions, manage campaigns and operate their systems.
The real opportunity is not simply to introduce another tool. It’s to develop intelligent capabilities that become a useful part of how the business works.
I help B2B organisations explore, design and engineer AI-enabled workflows, agents and connected applications around practical marketing and revenue operations requirements.
From account intelligence and campaign operations to CRM workflows, knowledge systems and automation across multiple platforms, I combine commercial operations experience with software and AI engineering to turn ideas into working capabilities.
Whether you have a specific use case in mind or want to understand where intelligent automation could make a difference, I can help you develop the right approach.
What could we engineer?
15 ways to apply intelligent engineering across four areas of commercial work.
What could we engineer?
Intelligent capabilities built around the work your team needs to do.
Every organisation has different systems, processes and opportunities.
Some applications can extend existing automation. Others require an AI model to interpret information, make recommendations or coordinate work across connected tools.
Here are examples of what could be designed and developed.
These are examples of intelligent capabilities I can design and engineer around a team’s systems, data and working requirements. They draw on professional experience and practical engineering work; the right implementation is defined for each organisation.
Show applications for
Fifteen applications in four groups.
Campaigns, activation and nurture
From campaign briefs to more connected execution.
Campaign operations involve a considerable amount of preparation, coordination, checking and follow-through.
Intelligent assistants and agents can support different parts of that operation.
- PlanCampaign planning and preparation
- AudienceAudience and segmentation support
- Build and checkCampaign setup and quality assurance
- LaunchResponse management and routing
- NurtureNurture and follow-up orchestration
- LearnPerformance analysis and improvement
Plan
Campaign planning and preparation
“We agree the brief, then lose a week turning it into tasks, lists and settings.”
Turn an agreed campaign brief into structured requirements, implementation tasks, audience needs, dependencies and operational checklists.
Where a person decidesYour team owns the strategy, the brief and the plan it signs off.
The practice behind itDemand Generation, ABM & GTM Operations
Audience
Audience and segmentation support
“Every list is rebuilt by hand, and nobody is sure who was left out, or why.”
Prepare and validate audiences using agreed targeting criteria, customer information, eligibility rules and campaign requirements.
Where a person decidesYou approve the definitions and the final audience.
Related engineeringOperations Control Tower
The practice behind itDemand Generation, ABM & GTM Operations
Build and check
Campaign setup and quality assurance
“Launch-day checks depend on who has time, and mistakes are found by recipients.”
Help teams configure workflows, review campaign assets and settings, identify inconsistencies and prepare campaigns for launch.
Where a person decidesNothing launches until a person releases it.
Related engineeringCampaign QA and operational assurance
The practice behind itDemand Generation, ABM & GTM Operations
Launch
Response management and routing
“Responses arrive in five places, and the good ones wait days for the right person.”
Interpret incoming enquiries, organise information, identify relevant follow-up requirements and connect responses with the appropriate teams and systems.
Where a person decidesRules decide the routing; a person approves each handover that matters.
Related engineeringAI-assisted enquiry handling in HubSpot
Nurture
Nurture and follow-up orchestration
“Our journeys were set up two years ago, and nobody knows who is in which, or when they leave.”
Coordinate lifecycle journeys, identify relevant next steps, prepare follow-up activity and help teams manage customer and prospect communication more consistently.
Where a person decidesPeople approve the journey, the messages and every send.
Related engineeringOpportunity signals and follow-up orchestration
The practice behind itDemand Generation, ABM & GTM Operations
Learn
Performance analysis and improvement
“The monthly review is two days of exports and an argument about whose numbers are right.”
Bring together campaign and operational information, prepare useful analysis and help teams identify where further investigation or improvement would be valuable.
Where a person decidesInterpretation, and what to do next, stays with your team.
The practice behind itRevenue Systems & Marketing Operations
These applications can be developed individually or connected into a broader campaign operating capability, depending on the organisation's needs.
CRM and revenue operations
Make commercial systems more intelligent and responsive.
AI can support how organisations understand and operate the information, processes and workflows inside their revenue systems.
Potential applications include:
- CRM diagnostic assistants that help investigate data, workflows and operational behaviour.
- Intelligent enquiry classification, lead handling and routing.
- Recommendations and workflow actions informed by customer or account information.
- Assistance with documenting, analysing and improving operational processes.
- Connected workflows that bring information and actions together across commercial platforms.
The aim is to make the systems more useful to the teams relying on them, not simply add another automation layer.
CRM diagnostics
“We know the CRM is wrong somewhere, but not where, or what a fix would break.”
Read-only inspection of data, lifecycle, routing and workflows, each finding marked observed, inferred or not established, and a repair rehearsed and priced before anyone makes it.
Where a person decidesYou choose which repairs to make, and approve each change.
Related engineeringUnderstanding HubSpot revenue flows before making changesTrigebra: a HubSpot AI-readiness diagnostic
The practice behind itHubSpot Engineering
Approved actions in the CRM
“Account signals turn up in Slack and inboxes, and most never reach the CRM.”
A recommendation with its evidence, the owner’s approval, then one controlled action in the CRM: once per signal, never without approval, and stoppable at any moment.
Where a person decidesThe account owner approves every action.
Clean-up and change planning
“Hundreds of properties and workflows, half unused, and nobody dares remove one.”
An inventory of properties, workflows and integrations with what depends on each, proposed consolidation, and a staged change plan with a way back at every step.
Where a person decidesYou approve every change; the assistant only proposes.
The practice behind itHubSpot Engineering
GTM intelligence and research
Turn available information into useful commercial intelligence.
Research and account intelligence often involve gathering information from different places, interpreting what matters and deciding what deserves attention.
I can help engineer capabilities that support:
- Researching companies, contacts, buying groups and markets.
- Monitoring relevant developments and account activity.
- Preparing structured, evidence-backed account briefs.
- Identifying changes or opportunities that warrant further investigation.
- Connecting research findings with existing GTM processes.
These capabilities can support more informed decisions while reducing the repetitive work involved in gathering and organising information.
Account research briefs
“Everyone researches accounts their own way, and nobody can check what a brief rests on.”
Research from public sources into a fixed brief, every quote checked against its source in code, what was observed kept apart from what was inferred, and what could not be found said plainly.
Where a person decidesA person decides whether to act; the brief advises.
Related engineeringAccount intelligence with verifiable evidence
Market and account watching
“We hear about the funding round or the new hire weeks after it mattered.”
Scheduled collection from job boards, press and newsrooms, deduplicated and classified, each run saying which sources answered in full, and only what merits a look passed on.
Where a person decidesYour team decides which signals matter, and what happens next.
Target-account evidence
“Sales and marketing disagree on which accounts are worth the effort, and neither can show why.”
Account lists built from evidence you can inspect: each company identified correctly across sources, and the reason for every inclusion and exclusion recorded.
Where a person decidesYou set the criteria and approve the list.
Related engineeringAccount intelligence with verifiable evidence
The practice behind itDemand Generation, ABM & GTM Operations
Knowledge systems, monitoring and connected agents
Develop new ways for people, information and systems to work together.
AI applications can also support broader operational needs.
Examples include knowledge assistants that help teams find relevant information, monitoring systems that identify exceptions, and connected agents that coordinate tasks across approved tools and platforms.
The appropriate architecture depends on what the application needs to accomplish.
Sometimes a straightforward automation is sufficient. Other requirements benefit from AI interpretation, specialised tools or an agent capable of carrying out a defined sequence of work.
Answers from approved knowledge
“The same questions about products, policy and process get a different answer from every person.”
An assistant that answers only from approved content, shows where each answer comes from, and hands over to a person when it cannot answer with confidence.
Where a person decidesYou approve what it may draw on, and who it hands over to.
Related engineeringThe assistant on this site
Systems and data-quality monitoring
“We find out a sync broke or a workflow stopped when the month-end numbers look odd.”
Read-only checks on workflows, integrations, scheduled jobs and data quality, run on a schedule, with a watchdog on the checks themselves and an alert only when a person is needed.
Where a person decidesPeople decide what is fixed, and when.
Related engineeringOperations Control TowerScheduled cloud research and monitoring
Connected, tool-using agents
“The work crosses four tools, and someone copies between them all day.”
Agents that plan and carry out a defined multi-step task through narrowly permitted tools and APIs, with every consequential step gated, logged and stoppable.
Where a person decidesYou decide what each agent may touch, and approve what it may not do alone.
From an AI idea to a working capability
Understanding the operation comes before choosing the technology.
A useful AI application needs more than a model and a prompt.
It needs a clear purpose, access to the right information, connections to the relevant systems and a practical way to operate.
My approach combines business understanding with hands-on engineering.
Explore the opportunity
Understand the work, the people involved, the information available and what the application should help accomplish.
Design the solution
Determine the appropriate combination of software, automation, AI models, agents, integrations and operational processes.
Build and connect
Develop the application and connect it with the relevant tools, data and workflows.
Evaluate and improve
Test how it performs against realistic requirements, refine its behaviour and make sure the surrounding process is dependable.
Prepare for operation
Document the system, establish appropriate responsibilities and support handover, ongoing improvement or further development.
The work can begin with a focused feasibility exercise, a prototype, a defined implementation or specialist engineering support within a larger initiative.
Engineering intelligent systems responsibly
Useful intelligence needs dependable engineering.
AI can interpret information and support decisions in ways that conventional automation cannot always achieve.
But an intelligent application also needs reliable connections, accurate data handling and a clear understanding of what it should do when information is incomplete or something unexpected happens.
I approach these systems by combining AI capabilities with software engineering, practical evaluation and appropriate operational controls. That includes considering:
Information and context
What the application needs to know, where the information comes from and how it should be interpreted.
Actions and integrations
Which systems it needs to interact with and what work it should be able to carry out.
Decision-making and accountability
Which decisions can be automated, which need human judgement and how the application should respond to uncertainty.
Testing and improvement
How to evaluate behaviour, identify problems and improve the system as requirements evolve.
The level of control should reflect the work being performed. Not every useful automation needs manual approval at every step.
Engineering in practice
From an account signal to a useful next step
When something relevant changes at an account, AI can help interpret the information and prepare a recommendation. Software checks the rules and avoids duplicate work. The responsible person decides whether to act, and the approved follow-up is recorded in the CRM. The value is a more dependable way to turn information into coordinated action, not simply another alert.
Explore how the system works. The interactive demonstration lets you examine the different components, how they interact and how the overall workflow behaves under different conditions. Switch between the business and engineering perspectives to explore the operational purpose and the technical design.
See the connected workflow in the LabExplore the AI, business rules, approval step and CRM connection behind this engineering example.The job
“When something happens at an account we care about, the owner should hear about it with a reason, and if they agree, get one task in the CRM. Never without their say, never twice, and stoppable at any moment.”
Done means: One task, for the right owner, on the right company, with the evidence attached.
Choose a ring, follow the run, or pick a failure.
How the workflow operates
Five rings, one action at the centre.
AI sits in an outer ring: it reads and proposes, and nothing more. Ordinary software checks what it proposes. A person approves. Only the centre writes to the CRM, and only once. Each ring is limited by what it can reach, not by what it is asked to do.
The model holds no tools and no keys. Every write passes a format check, a claim on the signal, a kill switch read and a bound, expiring approval, and is confirmed by reading the record back.
Choose a ring
Ring 1 of 5 · outside
The signalSignal intake
Something arrives from outside: here, a job post saying a company is rebuilding its lead routing. Whatever arrives from outside is treated as untrusted text.
An authenticated intake. The signal’s text, source link and the sentence that matters are kept with it, and passed on as data, never as instructions.
May
- Start one evaluation per signal.
May not
- Reach the CRM, a person or the model’s settings directly.
When it fails
A malformed or unauthenticated signal is refused before anything reads it.
Ring 2 of 5 · inside
AI reads and recommendsAI interpretation
An AI model reads the signal and recommends pursue, watch or reject, quoting the evidence it relied on and naming what it is unsure of. AI earns its place here: the signal is unstructured text that fixed rules judge badly. Only signals the rules let through ever reach it.
A small model called through its API, with a fixed answer format and no tools. An answer that does not fit the format goes no further. Instructions found inside the signal are treated as content to report, not orders to follow.
May
- Read one signal.
- Return one recommendation in a fixed format, with its evidence and its doubts. Only “pursue” goes on to a person.
May not
- Use tools.
- Read or write the CRM.
- Contact anyone.
- Approve anything.
When it fails
If the model is unavailable or its answer is unusable, nothing happens, and the run says why.
Ring 3 of 5 · inside
Software checksPolicy gate
Ordinary software, not AI, does the deciding that rules can do. Before any model call, it stops customers, duplicates, low-priority accounts and incomplete or ambiguous signals. After it, it checks the answer is well formed, that the signal has not already been acted on, and that the capability is switched on.
Stops before the AI (no model call); format validation after it; a claim table that lets exactly one approval own each signal; a kill switch read immediately before every write; spending capped at the model provider. The display score is set here, by rules, not by the model.
May
- Stop anything.
May not
- Approve anything, or change what the AI recommended.
When it fails
It fails closed: if a check cannot be completed, nothing is written.
Ring 4 of 5 · inside
A person decidesApproval
The account owner sees the signal, the evidence, the company and what would be written, and approves or rejects. An approval covers this one action, and the window it opens for the write closes after two minutes.
A signed link for this one run, showing the evidence and the CRM target; the reviewer’s typed name recorded; a write gate that expires two minutes after it opens.
May
- Approve or reject this one action.
May not
- Approve anything else.
When it fails
No approval, no action. If the write cannot happen inside its window, nothing is written.
Ring 5 of 5 · the centre
One action, onceGated write
The company is created or updated and one task is set for its owner; then the record is read back to confirm. Retrying a failed run finishes the job without doing it twice.
A key that cannot reach deals, owners or workflows; writes keyed to the signal; a done marker; confirmation by reading the record by its id; every write audited.
May
- Write the company’s signal fields and one task for this signal.
May not
- Touch deals, workflows or any other company.
When it fails
A crash leaves either nothing written or a state that a retry completes once.
Follow an example
Five steps, from signal to CRM follow-up.
Step 1 of 5
- Step 1 · The signalA change at an account is identified.
- Step 2 · AI reads and recommendsAI brings together the relevant information and recommends a next step.
- Step 3 · Software checksRules check eligibility, duplicates and the boundaries of the action.
- Step 4 · A person decidesA person reviews and approves the recommendation.
- Step 5 · One action, onceThe connected workflow records the approved follow-up in the CRM.
When something goes wrong
Seven failures, and where each is caught.
Six are handled, and tested. One is not yet: it is listed here because it is real.
Failure 1 of 7
The signal contains instructions
Caught by: AI reads and recommends
Text inside a signal tells the model to recommend “pursue”. The model treats it as content, quotes only genuine evidence, and names the planted instruction as a doubt. Nothing it says can act on its own.
Failure 2 of 7
The model is down, or its answer is unusable
Caught by: Software checks
An answer that does not fit the format, or no answer at all, goes no further. The run records why, and nothing is written.
Failure 3 of 7
Two approvals arrive at once
Caught by: Software checks
Both try to act on the same signal. One claims it; the other stops. One company, one task.
Failure 4 of 7
Someone switches it off
Caught by: Software checks
The switch is checked immediately before every write, so even an approval already given writes nothing. Switching on and retrying completes it once.
Failure 5 of 7
The CRM is slow to show a new record
Caught by: One action, once
The confirmation step could not find a company created seconds earlier. The run stopped instead of guessing. Confirming by record id fixed it; the retry made no duplicate.
Failure 6 of 7
It crashes just after approval
Caught by: One action, once
Either nothing has been written, or a retry finishes the job once.
Failure 7 of 7
A second signal about a company created seconds ago
Not yet caught. It would reach: One action, once
Duplicates are prevented per signal, not per company. Inside the CRM’s few seconds of search lag, a different signal for the same new company could create a second company.
Experience and engineering
Commercial understanding meets technical capability.
My background in B2B marketing and revenue operations informs how I approach intelligent systems.
I understand the processes, platforms and practical requirements behind CRM operations, campaigns, account intelligence, lifecycle marketing and connected GTM workflows. That experience is combined with ongoing engineering work across software, automation, integrations and AI-enabled applications.
Explore selected technical projects and working capabilities in the Lab, alongside the broader professional experience behind MarOps Lab.
A judgement, not a default
Sometimes the right amount of AI is none.
- No visitor text leaves the browser
The “Ask MarOps Lab” assistant on this site answers questions about the practice. A language model was evaluated for it, with the whole site as its knowledge.
It mostly kept to what the site can support. But it once invented a structure and attributed it to the site, partly obeyed an instruction planted in a question, and often spoke in the founder’s voice against its rules.
So the assistant uses no model. It matches questions against approved answers, and sends anyone it cannot answer confidently to a person. A seam is in place so that a model can sit above it later, with the matcher as the floor.
Working together
From exploring an opportunity to building the solution.
You might already know exactly which workflow you want to automate.
You might have an idea for an agent or intelligent application but need help understanding its feasibility, architecture and integration requirements.
Or you may be exploring how AI could extend the capabilities of an existing marketing or revenue operations team.
I can help you assess the opportunity, design an appropriate solution and undertake the engineering needed to bring it into operation.
The engagement should reflect the requirement, from a focused exploration to a defined build or ongoing specialist support.
Have an idea for an AI workflow, agent or intelligent application?
Whether you’re looking to improve an existing process, develop a new capability or connect AI with your commercial systems, we can start with what you want to accomplish.
You don’t need to arrive with the technical architecture already worked out.
Discuss a requirement