Jeff Brokaw

Your Team Can Build the AI Tool. Should They?

Jeff Brokaw · CMO and Certified Chief AI Officer ·

An AI coding assistant can make a software quote look ridiculous. Before you cancel the subscription, find out which parts of the job your team just volunteered to own.

I use AI to write code. It lets me build and change products without waiting for a traditional development cycle for every idea. I still have to decide what the product should do, check the result and deal with what breaks.

Pick a specific job and make the options compete on completing it. The custom build has to earn its budget too.

Should you build or buy AI marketing tools?

Buy when an existing product completes the required work at an acceptable cost and your team can operate it. Configure software you already own when it can handle the job without another platform. Build when a valuable requirement remains unmet and you can support the resulting system. Leave the process alone when changing it would cost more than the problem is worth.

Those choices can coexist. You can buy the commerce platform, use its existing recommendation features and write a small piece of custom logic for one placement. You do not have to replace the entire platform to control one decision.

My AI marketing operating model covers the people, evidence and permissions around the work once you choose how to run it.

What my personalization build actually demonstrates

My cold-start personalization demo starts with the same “shoes” example for two synthetic shoppers. One scenario supplies running intent. The other supplies wedding intent. Change that input and the product shelf changes.

Synthetic storefront after selecting wedding intent: the mixed catalog is reordered and each product shows a ranking reason.
Working synthetic demonstration. The mixed-catalog ranking changes; category eligibility and commercial outcomes require separate checks.

The implementation uses product scores and explicit weights. It exposes why an item moves. The ranking step does not need a language model to invent a recommendation on every request.

That is a useful build decision to inspect. The behavior can be checked against the inputs instead of accepting a plausible explanation from a model. A merchant can then ask whether the chosen products follow the commercial rules it intended.

The shoppers, catalog and store are synthetic. The demo demonstrates ranking mechanics. Its projected conversion and order-value figures are not measured customer results, and it does not prove access to real visitors' identities or search histories.

The demo reorders a mixed catalog, so an accessory can rank above a shoe. A production shoe-search placement would need category eligibility and stock checks before ranking. What signal does the business actually have when the page loads? May it use that signal for this purpose? What happens when it is missing?

I would give a vendor those same questions before commissioning a build. A custom interface does not establish that the underlying capability is missing from the market.

Check the software you already own before adding another tool

Suppose the immediate problem is that a Shopify product page recommends irrelevant accessories. Start by checking the catalog and the existing recommendation settings.

Shopify Search & Discovery is a free Shopify app for search, filters and product recommendations. Its recommendation documentation explains that generated related recommendations can use store purchase history, product descriptions or related collections. Merchants can also select related and complementary products manually.

That is store and product context. It does not require the arriving shopper to have bought from you before. A first-time visitor to an established store is a different problem from a new store with no sales history.

If the job is to show an appropriate accessory beside a product, configuration may be enough. Check the theme supports the relevant sections or blocks. Confirm the product relationships and availability. Measure what happens after the recommendation appears.

Do that work before someone prices a custom recommendation engine. The free app still takes staff time to configure and maintain, and the Shopify store itself has costs. You may nevertheless be much closer to an answer than the new software proposal suggests.

When does a paid personalization product deserve the budget?

When it handles a requirement you can name and demonstrate.

Rebuy advertises a starting price of $25 per month for one package, with pricing affected by the selected packages and monthly Shopify orders. That is a starting price for a defined package, not the whole platform at every volume. Get the quote for the functionality and order tier you will use.

Its URL and UTM rules are a useful example of why first-visit personalization is not automatically a custom-build project. The documentation describes using campaign parameters for merchandising when the relevant tracking consent is present. You can have useful campaign context before a shopper builds an on-site history.

That does not establish the visitor's identity or reveal everything they searched elsewhere. It gives the system particular inputs. Ask which inputs survive navigation, when they expire and what the rule does without them.

Rebuy's data-source documentation also describes product-based recommendations and warns that stores with little or no sales history should not expect meaningful variation immediately. Put the vendor's data requirements beside your actual data.

For a trial, ask the vendor to run your own scenarios: a campaign arrival, an unknown visitor, an unavailable product and a missing signal. Inspect the placement you intend to use. Require the quote to include the features needed to run those scenarios, rather than the lowest number on the pricing page.

Where n8n fits, and what you still have to build

n8n is an option for connecting a workflow around existing systems. It can be part of a custom solution; it is not a finished storefront recommendation product.

For this example, I would use a scheduled workflow to read approved catalog changes, prepare proposed recommendation rules and send them to a merchandising owner's review queue. An approved change could go to staging for checks before publication. The storefront would still need a way to serve the approved recommendations within its response-time budget.

The proposed workflow has a specific daily job for the human: inspect the changed items, the source data and the proposed rule before approving it. A failed catalog fetch should hold the update and alert the owner. A retry should not publish the same change twice. The previous approved configuration should remain available if the replacement fails.

n8n documents human approval for selected AI tool actions. That capability needs configuring. It does not make every action in a workflow approved by default.

There is also an ownership choice underneath the workflow. n8n Cloud handles infrastructure management. With self-hosting, your team takes that on. The free Community edition excludes some capabilities listed in the edition comparison, including SSO, environments and workflow or credential sharing. Price the edition your team needs.

For agencies, check the license FAQ against the service you intend to sell. It permits consulting and automation services, while restricting hosted arrangements that let clients build or modify n8n workflows. A free internal deployment is not blanket permission to resell the platform.

Compare the same job, not three unrelated feature lists

These options are starting points for investigation, not substitutes with identical scope.

Four ways to complete the same job. Scroll sideways on small screens.
Route What it could handle in this example What the team must establish
Configure Shopify Search & Discovery Related or complementary products, search settings and filters The existing features satisfy the requirement; the theme can display them; the catalog is suitable.
Buy an appropriate Rebuy package Paid recommendation and merchandising capabilities, including documented campaign rules The selected package supports the required placement, inputs, consent behavior and volume at the quoted price.
Assemble a custom solution, potentially using n8n A specific rule or cross-system process the existing products do not adequately handle Reliable data, serving behavior, review, monitoring, recovery and a capable ongoing owner.
Keep the current process Avoid changing a low-value or infrequent task The cost and consequence of leaving it alone are smaller than the cost and risk of replacing it.

Give each option the same input and judge the required output. If the proposed build only wins because it skips consent handling, stock checks or support, the comparison is unfinished.

For a lead workflow, the accepted output might be a correctly routed opportunity with the required context. For reporting, it might be a checked brief delivered before the spend meeting. My articles on AI lead qualification and automated marketing reporting cover those definitions. The purchase test should use the outcome the business needs.

What belongs in the cost comparison?

Use the additional cost of running the required workflow over the same period. I would start with twelve months, then inspect a higher-volume case and the cost of leaving.

Count setup and migration, the relevant subscription or license, hosting, data access, model calls where needed, routine review, maintenance and recovery work. Include any extra plan required for the access controls or collaboration your team needs. Keep quoted costs separate from estimates and measured costs.

Existing software is not automatically free to extend. Custom code is not automatically cheap because the first version took an afternoon. Both may consume the same employee who already owns important work.

Record cash spending and staff time separately. If you value that time to compare alternatives, state the hourly assumption and the work it would displace. Do not describe freed capacity as a cash saving unless an actual expense goes away. Avoid counting the same labor cost twice.

Cost per attempted run can look attractive while reviewers keep rejecting the output. Track the cost and human effort per accepted task, including correction and retry work. Only compare acceptance rates across equivalent tasks and standards.

Do not fund the full build because a vendor's starting price looks high. Ask for the relevant quote, complete the comparable scope and make the calculation. The marketing stack review applies that discipline to renewals; this is the decision before you add or replace a capability.

Download the build-versus-buy worksheet. It has separate fields for scope, cash, staff time, evidence and ownership. Fill it with your records or quotes; it contains no invented vendor estimates.

What can you prove in a one-week test?

Whether the candidate can complete a bounded job reliably enough to justify a larger pilot. A week is not a promise of measurable revenue lift.

Day one: define the job and the current process. Specify the input, accepted output, allowed actions and person who approves a consequential change. Record what the existing process costs, what failure looks like and what the business loses by leaving the problem unsolved. Keep measured losses separate from expected benefits. Set acceptance limits before looking at the candidate's results.

Days two and three: test the cheapest plausible option first. Configure existing software where possible. Give a vendor and a prototype the same permitted inputs. Include missing data, conflicting instructions and an unavailable product. Start with synthetic or appropriately authorized test data. Do not buy access to personal data simply to make a demonstration look complete.

Days four and five: count accepted work and the effort around it. Record task completion, corrections, human minutes, processing time and failures. Keep a count of cases the system correctly held for review. Have the responsible person try to reject or reverse a change. A human-review button that cannot stop publication has failed the test.

Day six: hand it to the next operator. Give someone other than the builder the run instructions. Can they identify a failed connection, pause the workflow and restore the previous approved version? What support would they need? Include that work in the ownership estimate.

Day seven: decide what earns the next investment. Buy, configure, fund a narrow build, change the requirement or stop. Record the evidence and the next review date in the marketing decision record. If the next question is commercial lift, design a suitable control and allow the outcome time to mature.

Who should approve the build?

The business owner should approve the required outcome and the investment. The team operating the workflow should accept the daily responsibility. Technical, security and relevant data owners should confirm the dependencies and access the proposed system requires.

Before a customer-facing change goes live, somebody needs the authority to reject it and restore the last approved version. My human-review guide covers that gate, and the vendor review covers the data path behind the product.

An agency proposal should show which existing features were tested, which requirement remains unmet and what the custom work adds. Name who will maintain it and what the client can inspect. If configuration solves the problem, recommend it.

If every run needs the builder, put their time in the budget.

Frequently asked questions

Is it cheaper to build or buy an AI marketing tool?

Compare the cost of completing the same required work over the same period. Include setup, the relevant subscription, data or model usage, hosting, review, maintenance and recovery. Record staff time separately from cash. A lower prototype cost or advertised starting price does not establish the lower ongoing cost.

Does building an AI tool mean training your own model?

No. A custom tool can combine existing software, model APIs and ordinary code. Some decisions can use explicit rules or scoring without a model call. The business still needs to test the output, control access and name an operating owner.

Can existing tools personalize a first website visit?

They can use inputs available during that visit, subject to the tool's features and applicable permissions. Rebuy documents campaign UTM rules, and Shopify documents recommendations based on product and store context. Neither example proves that an unknown visitor's identity or off-site search history is available.

Can n8n replace a paid marketing platform?

It may handle a specific workflow around existing systems. It does not automatically replace the platform's interface, recommendation logic, data infrastructure, support or controls. Define the required work, check the relevant n8n edition and license, and compare the complete operating responsibility.

What should an executive ask before approving a custom AI build?

Ask which valuable requirement existing products cannot adequately meet, what evidence supports that conclusion, who will operate the build, what the full cost includes and how the team will stop or recover a failed action. Require a bounded test against the current process before expanding the investment.

How long should a business test an AI marketing tool before judging ROI?

A one-week test can expose task failures, review burden and integration gaps. Commercial ROI needs a suitable comparison, enough relevant observations and time for the business outcome to mature. The required duration depends on the purchase cycle and test design, not a universal seven-day rule.

Sources