Why better data governance is the key to better AI

Arti­fi­cial intel­li­gence is every­where. If a busi­ness isn’t using AI, then it’s either claim­ing to use it or claim­ing that it’s about to start any day now. What­ev­er prob­lem your com­pa­ny is hav­ing, it seems that a solu­tion pow­ered by deci­sion intel­li­gence, machine learn­ing or some oth­er form of AI is avail­able. Yet, beneath the mar­ket­ing hype, the truth is that many busi­ness­es can indeed ben­e­fit from this tech – if they take the time to learn what it can (and can’t) do for them and under­stand the poten­tial pit­falls.

In essence, AI enables its users to do use­ful things with a large pool of data – for instance, fish out insights with­out tying up the time of data sci­en­tists. Data is there­fore fun­da­men­tal to AI. There is a direct rela­tion­ship between the qual­i­ty (and quan­ti­ty) of what’s fed into a machine-learn­ing appli­ca­tion and the accu­ra­cy of its out­put.

Data gov­er­nance has tra­di­tion­al­ly been viewed in terms of com­ply­ing with reg­u­la­tions that stip­u­late how data must be col­lect­ed, stored and processed. But AI has intro­duced new chal­lenges and risks to be man­aged. It’s not enough to obtain a vast amount of data; you also need to con­sid­er its char­ac­ter­is­tics. Where is it com­ing from? What does it actu­al­ly rep­re­sent? Is there any­thing that you need to account for before feed­ing this mate­r­i­al into your algo­rithm? Will it train the algo­rithm in the right things?

“We can use AI to iden­ti­fy unusu­al pat­terns of behav­iour in a busi­ness… or we can see a busi­ness chang­ing how it earns mon­ey from con­tracts in real time,” says Fran­ki Hack­ett, head of audit and ethics at data ana­lyt­ics firm Engine B. “To do this, you obvi­ous­ly need a clear idea of what is rel­e­vant, along with high-qual­i­ty gov­er­nance process­es over your AI. Oth­er­wise, you find either that there are far too many ‘risky’ items to con­sid­er or that the AI points you in the wrong direc­tion.”

More input required

One way of approach­ing data gov­er­nance is to use a tool known as an observ­abil­i­ty pipeline. This ensures that every process is vis­i­ble, col­lect­ing data that is then uni­fied and cleaned up to cre­ate a more con­sum­able final data set.

An exam­ple would be the con­ver­sion of raw web­site logs to an ana­lyt­ics plat­form. The orig­i­nal data and its point of con­sump­tion are ‘buffered’ by the pipeline: the raw data enters it and is processed before being sent out to where it needs to be con­sumed. The method of con­sump­tion can eas­i­ly be altered because the under­ly­ing data is unaf­fect­ed – that is, you can change how the data is pre­sent­ed with­out chang­ing the col­lec­tion process.

AI can both ben­e­fit from this process and become part of it. The pipeline itself may feed an algo­rithm, but machine learn­ing can be used to detect anom­alous data (based on past trends) before it gets too far. This can save peo­ple the time and effort they would oth­er­wise need to spend on check­ing and clean­ing data and, once it’s been processed, inves­ti­gat­ing any irreg­u­lar­i­ties. But it can also ensure that busi­ness-crit­i­cal algo­rithms aren’t being fed data that will lead them to draw the wrong con­clu­sions and, poten­tial­ly, nul­li­fy any ben­e­fits gained from intro­duc­ing AI to the process in the first place.

Ensur­ing observ­abil­i­ty has plen­ty of ben­e­fits for data flows that don’t involve AI, but the sheer vol­ume of mate­r­i­al involved and the com­plex­i­ty of machine-learn­ing process­es mean that it’s vital to know what’s hap­pen­ing to the data being processed. Check­ing that the num­ber of vis­i­tors in your web ana­lyt­ics match­es what your logs tell you is triv­ial com­pared with under­stand­ing the out­put of a com­plex algo­rithm that’s being trained and tuned over time.

This is because a sys­tem that might have start­ed by pro­vid­ing the insights you were seek­ing could be drift­ing ever fur­ther away from gen­er­at­ing any­thing use­ful. The bet­ter your view of what’s hap­pen­ing to the data, the eas­i­er it is for you to pre­vent this out­come.

The risks in this respect can be more seri­ous than, say, the poten­tial over­state­ment of a set of pro­ject­ed sales fig­ures. Dr Leslie Kan­than, co-founder and CEO of AI firm Tur­in­Tech, offers an exam­ple of where the stakes are much high­er: “If AI is applied to a hospital’s mag­net­ic res­o­nance imag­ing scans and it mis­di­ag­noses a seri­ous dis­ease such as pul­monary fibro­sis as bron­chi­tis, caus­ing the patient to take incor­rect med­ica­tion and expe­ri­ence adverse side effects, who is to be held legal­ly account­able?” 

He con­tin­ues: “Sim­i­lar­ly, if a data set and mod­el is con­sid­ered too accu­rate in its score, it could lead to over-rep­re­sen­ta­tion, mak­ing things go hor­ri­bly wrong. For exam­ple, an AI mod­el that is used to pre­dict future crim­i­nal behav­iour could over­fit data and incor­rect­ly come up with a bias against eth­nic minori­ties.”

Data gov­er­nance is key to ensur­ing that AI pro­duces use­ful results. It incor­po­rates an under­stand­ing of not only eth­i­cal and legal issues but the impli­ca­tions these have for what mate­r­i­al must be col­lect­ed and its poten­tial lim­i­ta­tions.

The organ­i­sa­tions that will ben­e­fit the most from AI will be those that take the time to build a frame­work that ensures they’re tar­get­ing the right data; col­lect­ing enough of it; check­ing and clean­ing it to ensure that it’s of a high stan­dard; and then using it in an appro­pri­ate, eth­i­cal way. 

With the right data gov­er­nance in place, these enter­pris­es can max­imise the ben­e­fits and min­imise the risks of using AI to pro­vide insights that will stream­line their process­es, inform their deci­sion-mak­ing and cre­ate pow­er­ful new prod­ucts and ser­vices. There is a lot more than hype behind what AI can do for your busi­ness – as long as you lay the right foun­da­tions for it.