The Modern Tech Stack: Data, Talent, Compute, and a Good Sales Team

Ask most companies about their tech stack and you get a list of software. Tanner Ingalls, who owns Gamify Data, put it differently on X: "Data. Talent. Compute. And a good sales team. This is the modern tech stack."

That line has no vendor names in it, and that is the point. In his view, software is the easy part to buy. What actually decides whether machine learning pays off is four layers: the history you own, the people who turn it into decisions, the compute to keep models current, and a sales team that acts on what the models say.

The cost of getting this wrong shows up in surveys. In a Gartner survey of 644 respondents, the top barrier to AI adoption, cited by 49% of participants, was the difficulty of estimating and demonstrating the value of AI projects.^[1] Our read: value goes missing when one of the four layers is missing.

Data: the history nobody else has

Your CRM, billing, product usage, and support records describe how your business actually ran. Which leads became customers. Which customers expanded. Which ones left, and what they did the month before. A competitor can buy the same software you run. They cannot buy that history.

It only counts as an asset if it is connected and recorded well. Salesforce's 2026 State of Sales research found that over half of sales leaders with AI (51%) say disconnected systems are slowing down their AI initiatives.^[2] The fixes are usually unglamorous: close lost deals instead of leaving them open, date every churn event, record the leads nobody worked. (Why the history is the asset: Your Company Data Is the Most Undervalued Asset in Your Business. How to record it: Data Collection That Improves Models.)

When it is missing: talent and compute have nothing specific to learn from. You end up with a generic score that could belong to any company, so it fits yours poorly.

Talent: people who build models and translate them

The talent layer is two jobs. One is building: turning raw records into inputs a model can use, and testing the model on periods it never saw. The other is translating: deciding what a score means for this week's call list, which threshold triggers a play, and who owns it.

The market for the first job is tight and getting tighter. The U.S. Bureau of Labor Statistics projects employment of data scientists to grow 35 percent from 2025 to 2035, much faster than the average for all occupations, and reports a median annual wage of $120,230 in May 2025.^[3] BLS also lists communication as a core skill: data scientists must be able to convey the results of their analysis to technical and nontechnical audiences to make business recommendations.^[3]

In our experience, translation is the scarcer half. A model can be accurate and still useless if nobody connects it to a decision. (The build side: Feature Engineering for Tabular Business Data and Evaluating Models for Real Decisions. The translation side: Churn Prediction: Finding Actionable Drivers.)

When it is missing: the data sits in the warehouse feeding reports. Or a vendor model scores your accounts and nobody in the building can say why, or when to stop trusting it.

Compute: abundant enough to keep models current

Compute used to be a reason to retrain once a year. That excuse is fading. Stanford's 2025 AI Index reports that at the hardware level, costs have declined by 30% annually, while energy efficiency has improved by 40% each year.^[4]

For the models most mid-market companies need, a churn score or a lead score on CRM and billing data, we find that compute is rarely the real limit. Habit is. Retraining monthly and scoring nightly is practical on the warehouse and cloud tools you likely already pay for. (How to do it without a new platform: Production ML Without Massive New Infrastructure.)

When it is missing: models go stale. Last year's buyers set this year's priorities, and the drift stays hidden until a forecast misses badly. (Why that happens: Why Forecasting Models Degrade.)

A good sales team: where a score becomes revenue

A model doesn't close deals. People do.

Rep time is the constraint. In the same Salesforce research, the average seller spends 40% of their time selling.^[2] That leaves a small window each week for real customer conversations, and lead prioritization is a claim about how to spend it. The claim only pays off if reps actually work the list in order.

In our view, that makes adoption a sales management job, not a data science job. Managers set the expectation that the ranked queue gets worked first. Comp and coaching don't reward cherry-picking around it. Reps get a quick way to flag a score that looks wrong, and those flags go back to the people building the model. The same goes for CSMs and churn lists.

A good sales team also feeds the data layer. Every logged outcome, a meeting booked, a deal lost with a real reason, becomes training data for the next model.

When it is missing: the scores land in the CRM and get ignored. Reps call who they always call, and six months later someone concludes "the model didn't work" when it was never tried.

A model doesn't close deals. Reps working the list do.

What breaks when one layer is missing

Our shorthand for diagnosing a stalled ML program:

• Data without talent: reporting, not prediction.
• Talent without data: clever models of the wrong thing.
• Both without compute discipline: a model that was right last year.
• All three without the sales team: an accurate score nobody uses.

Each layer limits the others. Spending more on the strongest one does not make up for the weakest.

How the layers connect in lead and churn work

Take lead optimization. Data: two years of inbound leads with dated outcomes. Talent: a model that ranks them, plus routing rules a sales leader agreed to. Compute: nightly scoring written into the CRM. Sales team: reps work the queue in order, a random holdout gets the usual treatment, and outcomes get logged. Those outcomes become next month's training data.

Churn works the same way, with CSMs in place of reps and save plays in place of calls. In both cases the loop only closes when the people at the end act on the score and record what happened.

What to ask for

1. An inventory of your proprietary history, with the gaps named.
2. A named translator who owns turning each score into a play.
3. A written schedule for retraining and scoring.
4. A sales commitment: reps work the scored list in order for a test period, against a random holdout.
5. A feedback field where reps and CSMs flag scores that look wrong.
6. One revenue metric all four layers answer to.

If you want lead and churn models built on the history you own and worked by the team you have, talk to Gamify Data. We bring the talent and the compute plan, and we design around how your reps and CSMs already sell.

Sources

1. Gartner, "Gartner Survey Finds Generative AI Is Now the Most Frequently Deployed AI Solution in Organizations," press release, May 7, 2024. https://www.gartner.com/en/newsroom/press-releases/2024-05-07-gartner-survey-finds-generative-ai-is-now-the-most-frequently-deployed-ai-solution-in-organizations
2. Salesforce, "The Productivity Gap: New Survey Shows 9 in 10 Sellers Are Betting on AI and Agents To Help" (State of Sales report, 2026), Salesforce News, February 3, 2026. https://www.salesforce.com/news/stories/state-of-sales-report-announcement-2026/
3. U.S. Bureau of Labor Statistics, "Data Scientists," Occupational Outlook Handbook, last modified August 27, 2026, accessed October 11, 2026. https://www.bls.gov/ooh/math/data-scientists.htm
4. Stanford Institute for Human-Centered Artificial Intelligence, "The 2025 AI Index Report," Stanford HAI, 2025, accessed October 11, 2026. https://hai.stanford.edu/ai-index/2025-ai-index-report