One of them says 80% of companies will fail at the thing everyone is rushing to build.

Another says data engineers will stop writing pipeline code by the end of 2027.

I watched every minute of Ataccama FWRD 2026. Here's what stood out, what got pushback, and the demo that made me rethink how we score data.

Watch the full sessions free: ataccama.com/fwrd-2026/on-demand

The number that set the tone

67% of data leaders say they don't fully trust the data in their own organization, even as they're asked to deliver trusted data foundations for AI.

Think about that. 2 out of 3 leaders are being asked to build for AI on data they don't fully trust. 

That gap drove the whole day.

Rewind: 3 teams who already did the hard part

Session 1 looked back at 2026.

Lennar. Isaac Gabay, VP of Data & AI, shared Lennar's AI platform journey and how his team builds trust in the AI agents it runs in production. His goal was simple. When an agent gives an answer, his team wants to be able to say the data behind it is certified.

Lloyds Banking Group. Jo Riseborough described how Lloyds approached data modernization, grounded in her belief that AI's success depends on the data discipline behind it.

iA Financial Group. Chantale Boulanger and Nickolaos Mikedis explained how they built a master data management foundation and a 360-degree client view, giving growth teams one shared picture of each customer.

All 3 stories are in Session 1: ataccama.com/fwrd-2026/on-demand

Fast FWRD: 5 predictions, 1 rule

Session 2 was my favorite part. Each panelist brought 1 prediction for 2027. Then Peter Stojanovic of HotTopics had the others stress test it. 

Here's every bet.

Prediction 1: Fix ownership before you build agents

"By the end of 2027, the enterprises with the most operational AI agents will be the ones that have moved data ownership and stewardship into the business first, before introducing AI."  Camilla Björkqvist, Mondelēz International

Her reasoning hit home. Most agents today are built by AI engineers in an experimental phase. Governance shows up late. When those agents go live, the gaps in ownership and quality show up fast.

Mondelēz made the call to settle ownership before it started building agents.

She also made a sharp point about how people and agents differ. A human working with messy data knows how to work around the gaps. An agent doesn't. So data owners need to define the boundaries: what an agent may do, and where a human must stay in the loop.

And without solid cataloging, she said, AI tools start guessing.

Prediction 2: Only 20% will get semantics right

"80/20 rule. In two years, only 20% of organizations will be succeeding with semantics."  Juan Sequeda, ServiceNow

The 20% will treat semantics with the respect it deserves and build strong foundations. The other 80% will wait for a magic wand, expecting AI to build the glossary and semantic model for them.

He walked through the history. First we had data swamps. Then dashboard swamps. Now he sees semantic layer swamps forming, with thousands of agents giving thousands of different answers to the same question.

His fix was maybe the least exciting advice of the day but the hardest truth. “It's like going to the gym and eating your vegetables”.

He compared it to the early cloud years. The trailblazers who did the foundation work won. Everyone else came back later and did the work anyway.

Prediction 3: Data engineers stop writing pipeline code

"The majority of pipeline work is going to get handed off to AI agents; zero human time spent writing code in the pipelines, debugging and managing them. The role of the data engineer is going to be about interfacing with the business teams."  Christopher Child, Snowflake

Zero. By the end of 2027.

So what's left for the data engineer? Chris said the key skill becomes understanding the business. Sit down with the VP of sales. Learn what they're trying to accomplish. Translate that into what agents need to do.

Camilla agreed and added a line from a former boss that I'm stealing: “Data problems are just people problems in disguise.”

She expects stewards to shift too. Less time checking flags, more time handling the exceptions machines can't.

Hear the full back and forth: ataccama.com/fwrd-2026/on-demand

Prediction 4: Governance runs itself

"By the end 2027, governance will move from AI-assisted to agent-driven."  Donovan Spronk, Deloitte

His logic: static policy documents can't keep pace with how fast AI is moving. Policies will turn into machine-verifiable data contracts.

He's working in Europe, and he sees the shift starting, though not at scale yet. He expects the EU AI Act to push governance up the agenda once the first incidents hit.

Juan backed him up with a practical point. The scope of what needs governing keeps growing: unstructured data, AI assets, even physical assets. The only way to cover all of it is to automate.

Prediction 5: The next AI stack won't come from the usual names

"By the end of 2027, we will see the emergence of modern AI stacks that may not come from the traditional hyperscalers."  Jay Limburn, Ataccama

He pointed to data platform companies expanding upward into the AI layer. His view: those platforms will only succeed if best-of-breed tools can plug into them, including through open standards like Apache OSI. 

He also made the case for master data management having a second act. MDM programs have a reputation for being slow and costly. But when agents need one trusted view of a customer or supplier, MDM suddenly sits right next to the AI outcomes.

As he put it, the problems we're solving today look a lot like the ones from 5, 10, 15 years ago.

The thumbs round

The session closed with a fast round of agree or disagree. Not everyone was on the same page. Chris gave at least 1 firm thumbs down.

These are the moments you only get from watching it: ataccama.com/fwrd-2026/on-demand

What's next: the roadmap session

Jessica Smith, Chief Product Officer at Ataccama, then gave a first look at where the Ataccama ONE platform is heading. 

3 things stood out.

1. MDM set up by agents, approved by people

Today an MDM project needs experts at every step: modeling, mapping, rules, each handing off to the next.

Jessica showed the new approach. Agents do the setup work across 6 steps. A person approves each one.

The screen in the demo shows the run in progress: 11 of 15 stages done, 36 minutes of agent work, $11.60 in agent cost, and 4 human decisions so far.

Anyone who has lived through a multi-year MDM program will feel that one.

2. Data quality checks that start with what agents use

Smart profiling looks at what AI agents actually use to find your critical data elements. Headless data quality opens every operation through MCP. In plain terms: quality checks can run inside the tools and agents your teams already use.

3. A trust score for each answer

This was the demo I keep thinking about.

A sales table scores 88.56% overall. Looks fine.

Then someone asks about EMEA bookings in Q1 2025. That slice scores 16.7%. The reason: a CRM migration gave 184 deals the wrong close date.

A table-level score would have hidden that completely. The agent would have answered with confidence. The answer would have been wrong.

Scoring the specific rows behind each answer is the kind of thing that makes agents safe to use.

My take

I cover a lot of conferences. This year every single one talked about agents.

FWRD was one of the few that spent 2 hours on what might be the most important question: what happens when an agent acts on bad data?

The 5 predictions point the same way. Settle ownership first. Do the semantic work. Let agents handle the repetitive parts. Keep people at the decision points.

None of it is flashy. Juan would call it eating your vegetables. But it’s what will set the data and AI leaders who put in the work apart from the rest.

I'm saving these 5 predictions. Ataccama says they'll check back at FWRD 2027 to see who called it. I'll be there to score them.

Which prediction do you agree with most? Which one do you think is wrong? Tell me below.

Watch all of FWRD 2026 free: ataccama.com/fwrd-2026/on-demand

Best,
Ravit Jain
Founder & Host of The Ravit Show