MarTech stack architecture: frameworks & methodology

The renewal notice lands 90 days out. Adobe wants a three-year commitment on Marketo Engage, priced against a contact database that has grown to 2.4 million records, and your demand gendemand genCreating and stimulating demand for your offer, often upstream of the buying process to generate interest and awareness before prospects are ready to buy.View full definition → lead has a counter-proposal: drop the automation platform, activate from the warehouse instead. One quarter to decide, and the decision is not really about Marketo. It is about which capabilities you are already paying for twice, what a switch costs across five years rather than year one, and whether your team can run a migration while still shipping campaigns.

The layer diagram, and which systems hold the record versus do the engaging, belong to the foundations lesson; assume them here. This lesson is the selection method: capability mapping, build-versus-buy scoring, total cost of ownership, and the order in which you sequence a replatform.

The stakes are measurable. Gartner's 2023 marketing technologymarketing technologyThe connected set of software tools a marketing team uses to plan, run, measure and automate campaigns across channels.View full definition → survey put stack utilisation at 33%, down from 58% in 2020, while martech holds roughly a quarter of the marketing budget. Scott Brinker's annual directory counted around 14,000 products in 2024. Two thirds of what you own is idle, and there are 14,000 ways to make that worse.

Capability mapping: the inventory that is not a tool list

Most stack audits start by listing tools, which produces a spreadsheet nobody can act on. Start from capabilities instead: verbs, written at the grain of a job somebody does weekly. "Suppress customers who filed a support ticket in the last seven days." "Trigger a message when a trial hits 80% of quota." "Attribute closed revenue to a paid channel within 24 hours of the CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → stage change."

A mid-market B2B team lands on 40 to 60 such capabilities. A group with hundreds of brands lands on several hundred. For each one, record five things: the team that owns it, every tool that claims it, the tool actually used, criticality (blocks revenue / degrades performance / convenience), and a quality score out of five from the person doing the work, not from the platform owner.

Two outputs matter. The first is duplicate coverage: capabilities where three tools claim the job and one does it badly. The second is the uncovered critical list, usually shorter than people expect and usually about identity, consent, or getting revenue data back out of the CRM.

Not every duplicate is waste, and this is where naive rationalisation destroys value. Running transactional email on a separate provider from marketing email is a deliberate duplicate: it isolates password resets and receipts from the reputation risk of a bad promotional send. A second analytics tool that finance trusts is a duplicate too, and cheaper than a quarter of arguing. Kill duplicates that nobody chose. Keep the ones somebody can defend in one sentence.

Scoring build, buy, or extend

Turn the decision into a weighted score before the demo, not after. Six criteria, weights summing to 100:

• Differentiation: does this capability change what a customer experiences in a way competitors cannot copy by buying the same tool? (25)
• Category maturity: are there four or more credible vendors, with published pricing and a live migration path out? (20)
• Integration cost: how many systems must exchange data, and does the vendor have working connectors for them or only an APIAPIApplication Programming Interface: a standardised interface that lets applications communicate and exchange data without knowing each other's internal workings.View full definition → and good intentions? (20)
• Switching cost: what does exit look like in year three, including data extraction and rebuild? (15)
• Internal capacity: is there a named engineer or admin with committed hours, permanently, not for the launch? (10)
• Compliance and data residencydata residencyThe requirement that data is physically stored and processed in a specific country or region, often driven by law or contract.View full definition →: does any market you sell in require the data to stay put? (10)

Build only when differentiation scores high and internal capacity is a real, funded commitment. The cost of building is not the build. Plan on 15 to 20% of the original build cost every year in maintenance, forever, and remember that the engineer who wrote it will leave.

The option that gets skipped is extend. Before any new purchase, read your own entitlements. Adobe bought Marketo in 2018 for 4.75 billion dollars specifically to fold B2B automation into a suite it already sold you, which means the capability you are about to procure may sit inside a licence you signed for a different reason. Same with the CRM. Vendors of suites (Adobe and Salesforce both sell the very thing this lesson is about) have a commercial interest in you not knowing what you own, and enterprise agreements are written to make counting hard. Make someone spend a day counting.

Total cost of ownership across five years

Licence price is the smallest interesting number. Model six lines over 60 months:

• Licence, with the annual uplift written into most enterprise contracts, plus the pricing metric. Marketo prices against database size, so every dead contact and every duplicate is a recurring line item. Deduplicate before you request the quote, not after you sign it.
• Implementation. For enterprise automation, first-year implementation commonly matches or exceeds the first-year licence.
• Integration engineering: build cost per connection plus permanent maintenance. Every integration is a liability that breaks when either side ships an update.
• Internal operations: 0.5 to 2 FTE of admin, template building, and QA.
• Training, plus retraining after turnover.
• Exit: data extraction, dual running, rebuild.

Run it with numbers. A 200k annual licence, 200k implementation, 1.5 FTE at 90k, 120k of integration build with 30k a year of upkeep, 25k a year of training: about 2.3 million over five years, of which the licence is roughly 44%. That ratio is the point. Two stacks with identical licence lines can differ by a million in ownership, and procurement negotiates only the smallest column.

Sequencing a replatform

Order matters more than tool choice, because a badly ordered migration costs pipeline while the good tool sits half configured.

1. Read the contracts before you read the RFPs. Termination windows, auto-renewal clauses, data export rights. The classic failure: signing the replacement, then finding the incumbent auto-renewed for 12 months, and paying for both.
2. Identity and consent move first. The persistent profile the foundations lesson describes, plus suppression and consent records, has to be correct in the new environment before any activation tool points at it. Migrating a suppression list badly means mailing people who opted out, which is a regulatory exposure under GDPRGDPREU regulation governing how organizations collect, store and use personal data, with fines tied to global revenue for breaches.View full definition →, not a hygiene issue.
3. Rebuild, do not lift and shift. Campaign logic lives in vendor-specific constructs, and an export gives you assets, not logic. Audit which programmes actually ran in the last 12 months; in most stacks the tail is enormous. Rebuild the programmes that carry volume, retire the rest, and write the retirement list down so nobody reinstates it from memory.
4. Parallel run one revenue-carrying programme end to end and reconcile counts against the CRM before touching a second one.
5. Warm the sending infrastructure. New domains and IPs need weeks of graduated volume. Send at full blast on day one and you land in spam folders across your entire base, and a damaged sender reputation costs more to repair than the licence you saved.
6. Decommission on a dated plan aligned to licence end, keeping a read-only archive for legal retention.

Budget six to twelve months of overlap where both platforms are paid for. That overlap is the single biggest surprise in most replatform business cases, and it is the reason a project justified on licence savings shows negative return in year one.

Unilever: mapping capabilities, not brands

Unilever runs roughly 400 brands across some 190 countries. If each brand picked its own tools, the group would hold thousands of contracts and no comparable customer record anywhere. Under Conny Braams, chief digital and marketing officer from 2019 to 2023, the direction was to pull first-party consumer data into shared platforms and take more media buying in-house, which only works if the unit of decision is the capability rather than the brand. Consent capture, identity, and measurement get decided once for everyone. Content production and channel execution stay local, because a Japanese team needs to run on messaging channels a global template will never cover.

The second-order effect is the one to plan for: centralising a capability removes a local team's ability to move quickly, and if the method does not include a sanctioned exception route with a time limit, those teams buy their own tools anyway and you rebuild the mess you just cleaned.

Siemens: sequencing against constraints you do not control

Siemens is organised into distinct businesses (Digital Industries, Smart Infrastructure, Mobility among them) with a large German workforce and customers in nearly every market. Two constraints reshape any sequencing plan there. In Germany, introducing systems that process personal data typically requires works council agreement, which adds weeks or months that no vendor timeline includes. And China's PIPL, in force since November 2021, pushes towards a separately hosted instance, which means two deployments, two integration sets, and a TCO model with the operations line doubled for that region.

Sequence by business unit and legal jurisdiction, and let the units with the fewest constraints go first so the pattern is proven before it hits the hard ones.

CMO action items

• Run the capability mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → before the budget cycle, at capability grain, with the quality score coming from the operator rather than the platform owner
• Ask legal for a single sheet listing every martech contract with its termination window and auto-renewal date, and keep it current; you cannot sequence a replatform without it
• Require a five-year TCO model and a written integration plan in every purchase paper above a set threshold, and reject papers that show licence cost alone

Common mistakes that kill results

• Buying a tool to solve a process problem: unclear ownership between sales and marketing survives any platform migration, and now it fails at higher volume and cost
• Treating migration as export and import, when campaign logic, scoring, and attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → definitions have to be rebuilt and re-agreed
• Justifying consolidation on licence savings alone, then absorbing the saved cost back into headcount because the surviving platform needs two admins instead of one
• Letting a suite vendor solve an architectural question by selling more of its own modules, which is sometimes the right answer and always the answer they will give